Paper deep dive
FlatLands: Generative Floormap Completion From a Single Egocentric View
Subhransu S. Bhattacharjee, Dylan Campbell, Rahul Shome
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/22/2026, 5:28:49 AM
Summary
FlatLands is a dataset and benchmark for single-view bird's-eye view (BEV) floor completion, designed to improve indoor navigation by inferring traversability maps from limited egocentric observations. It aggregates 270,575 observations from 17,656 real indoor scenes across six datasets, providing a rigorous testbed for uncertainty-aware generative modeling and monocular RGB-to-floormaps pipelines.
Entities (6)
Relation Signals (3)
FlatLands â contains â Matterport3D
confidence 100% · FlatLands combines six real indoor metric datasets: Matterport3D
FlatLands â usesforevaluation â ScanNet++
confidence 100% · ScanNet++ reserved strictly for test-only OOD evaluation
Subhransu S. Bhattacharjee â authored â FlatLands
confidence 95% · FlatLands: Generative Floormap Completion From a Single Egocentric View Subhransu S. Bhattacharjee
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:A single egocentric image typically captures only a small portion of the floor, yet a complete metric traversability map of the surroundings would better serve applications such as indoor navigation. We introduce FlatLands, a dataset and benchmark for single-view bird's-eye view (BEV) floor completion. The dataset contains 270,575 observations from 17,656 real metric indoor scenes drawn from six existing datasets, with aligned observation, visibility, validity, and ground-truth BEV maps, and the benchmark includes both in- and out-of-distribution evaluation protocols. We compare training-free approaches, deterministic models, ensembles, and stochastic generative models. Finally, we instantiate the task as an end-to-end monocular RGB-to-floormaps pipeline. FlatLands provides a rigorous testbed for uncertainty-aware indoor mapping and generative completion for embodied navigation.
Tags
Links
- Source: https://arxiv.org/abs/2603.16016v1
- Canonical: https://arxiv.org/abs/2603.16016v1
Trouble viewing inline? Open PDF directly â
Full Text
114,748 characters extracted from source content.
Expand or collapse full text
FlatLands: Generative Floormap Completion From a Single Egocentric View Subhransu S. Bhattacharjee â , Dylan Campbell, and Rahul Shome School of Computing, The Australian National University, Canberra, Australia Abstract.A single egocentric image typically captures only a small portion of the floor, yet a complete metric traversability map of the sur- roundings would better serve applications such as indoor navigation. We introduce FlatLands, a dataset and benchmark for single-view birdâs- eye view (BEV) floor completion. The dataset contains 270,575 obser- vations from 17,656 real metric indoor scenes drawn from six existing datasets, with aligned observation, visibility, validity, and ground-truth BEV maps, and the benchmark includes both in- and out-of-distribution evaluation protocols. We compare training-free approaches, deterministic models, ensembles, and stochastic generative models. Finally, we instan- tiate the task as an end-to-end monocular RGB-to-floormaps pipeline. FlatLands provides a rigorous testbed for uncertainty-aware indoor map- ping and generative completion for embodied navigation. Keywords:scene completion·embodied AI·generative modeling 1 Introduction Partial observability is a defining constraint in indoor autonomy â decisions must be made from sensory evidence that is incomplete, noisy, and inherently viewpoint-limited [2,5,35,61,104]. This paper targets a specific perception task that sits on this critical path âinferring a usable traversability map from limited observations. Robotic perception distills high-dimensional sensor streams into compact world models for mapping and planning under uncertainty [19,45,68,69], combining geometric representations [65,90,100] with probabilistic occupancy es- timates [41,97,144] and converting them into spatial abstractions such as metricâ topological [10,88,141,143] and semantic maps [67,107,118]. In contrast, we focus on a ground-plane birdâs-eye-view (BEV) floormap: a compact 2D grid of traversability obtained by projecting 3D geometry to the floor plane. BEV representations are widely adopted [73,166] and directly sup- port collision checking and reachability under uncertainty [85], making them an efficient abstraction for navigation tasks. Operational map-based reasoning from egocentric imagery has also shown strong practical value for localization and decision making [120]. Among sensing modalities, single-frame egocentric RGB is especially challenging [72,105,161]: one view observes a narrow frustum â Corresponding author:Subhransu.Bhattacharjee@anu.edu.au. arXiv:2603.16016v1 [cs.CV] 16 Mar 2026 2Bhattacharjee et al. Input Predictive StageGenerative StageFinal Output RGB frameI Depth mapD Floor segmentS BEV observed floorF obs BEV unobserved maskU Ί Image Segmentation Depth Estimation Κ Orthogonal Projection conditioning Camera Intrinsics Encoder Decoder Conditional model p prior Sampled BEV set Ë F (1) ,..., Ë F (K) p posterior Sampling K=4samples One-to-many inverse mapping Fig. 1:Pipeline. From a single RGB image, our model predicts depth and floor seg- mentation and projects them to BEV, producing observed floorF obs and unobserved maskU. A conditional generator then predicts floormap completions in the unobserved region, while preserving observed evidence. while most traversable floor is unobserved due to occlusions. BEV completion from a single image is therefore inherently ambiguous and belongs to the class of Bayesian inverse problems [133,138], where posterior reasoning is required rather than a single point estimate [33, 156]. Solving the single-frame case provides a per-step prior that can be fused temporally in multi-step navigation [4,51,110]. We studysingle-view BEV floormap completion: given one RGB image, infer adistributionover plausible traversability maps in the unobserved BEV while exactly reproducing observed labels. Since downstream decisions depend on un- certainty, we evaluate multi-hypothesis predictions using both fidelity and di- versity metrics [46]. Indoor layouts exhibit strong structural regularities (rooms, doors, corridors), which are often modeled via grammar and pattern-theoretic priors [67, 94, 108]. Modern generative models can internalize these regularities for posterior inference and sampling [52, 76], and are increasingly exploited in robotics under uncertainty [12,20,21,42,113,124]. Yet no standardized real-world benchmark exists for indoor BEVmetricfloormap completion. To this end, we introduce FlatLands, a dataset with 17,656 real indoor scenes from six sources [7,24,34,36,147,160]. Training data is synthesized from phys- ically feasible camera centers and yaw headings, with visibility computed by field-of-view checks and ray-based occlusion as shown in Fig. 2b. After automatic curation, the dataset contains 270,575 observations with aligned floor evidence, visibility, validity, and ground-truth BEV maps. Evaluation uses the full test split with explicit in-distribution (ID) and out-of-distribution (OOD) partition- ing, with ScanNet++ reserved strictly for test-only OOD evaluation and never used in training or validation. The task is also instantiated as a full monocular RGB-to-floormaps pipeline (Fig. 6). Quantitative evaluation uses standardized BEV-conditioned inputs, isolating completion quality from front-end estimation; end-to-end results confirm that method ranking transfers to realistic egocentric input. We benchmark training-free, deterministic, ensemble, and stochastic base- lines under a shared protocol (Secs. 5 and 5.2). Contributions.Our main contributions are as follows: Generative Single-View Floormap Completion3 1. We formalize the single-view indoor BEV floor completion task and re- leaseFlatLands, the first benchmark of its kind with 270,575 observations from 17,656 real metric scenes across six datasets of RGB images and BEV floormaps, each with visibility masks, validity regions where the dataset has a floor or non-floor label, and full provenance metadata. 2. We design an extensive evaluation suite covering masked multi-hypothesis calibration and a monocular end-to-end RGB-to-floormaps pipeline that demonstrates real-sensor viability. 3. We report the performance of eleven methods spanning training-free ap- proaches, deterministic predictors, epistemic ensembles, and three families of stochastic generators under a common conditioning input. We find that generative models better capture the observation-conditioned completion un- certainty than point estimates, and they also achieve stronger overall perfor- mance. A boundary-variance decomposition further reveals that epistemic ensembles conflate seed divergence with layout ambiguity, while conditional flow models localize uncertainty to structurally ambiguous regions. 2 Background and Related Work Indoor map completion and occupancy anticipation.Classical occupancy map- ping maintains per-cell Bayesian beliefs over traversability from sequential ob- servations [40, 41, 57, 96, 144], while variable-resolution occupancy formulations improve representation efficiency in large spaces [97]. Under multi-step exploration, OccAnt [110] trains a CNN anticipator on Habitat simulations to project observed evidence into a predicted occupancy map; Katyalet al. [64] and Katsumataet al. [63] exploit temporal sequences for spatial anticipation. Aydemiret al. [3] use relational priors over object co- occurrence to complete partially observed rooms. Methods such as FloorNet [78], FloorSP [25], and room-layout estimation from monocular imagery [92] aim to re- construct architectural plans from RGB-D, panoramic scans, or single-image cues with richer supervision. MapEx [51] and PIPE [4], which use the LaMa inpaint- ing network [135], target multi-modal uncertainty under scoring with multi-step observations, while [113] applies diffusion priors to 3D occupancy completion. In navigation, the robot primarily needs atraversabilitymap to denote where it can safely navigate [12,41,98]. Image inpainting and outpainting via generative models.Classical inpainting methods propagate structure into missing regions via PDE-based diffusion and exemplar matching [11,23,139]. Modern approaches are predominantly learned e.g. LaMa [135] couples large receptive fields (Fourier convolutions [31]) with adversarial training, while partial convolutions [79] build on the U-Net back- bone [117] by masking convolutional updates. Related ideas also appear beyond natural images, including occupancy inpainting for 2D grid maps [150]. Out- painting (image extrapolation) extends contentbeyondobserved boundaries and is typically less constrained than interior-hole inpainting, since boundary condi- tions are available only along the crop edge [149,159]. Recent learned outpainting 4Bhattacharjee et al. methods explicitly target this regime, emphasizing semantic consistency and di- versity [30,71]. A complementary line of work treats in and outpainting asconditional gen- erationunder missing evidence. Context encoders [99] introduced learned con- text completion, while diffusion-based inpainting methods such as RePaint [86] and broader pixel-space diffusion models [52,55,95,131] enforce conditioning via per-step masking. This masked-conditioning interface has since become stan- dard in large-scale generators that support inpainting and outpainting [15,101, 106,112,116]. For explicitly multi-modal completions, posterior-sampling mecha- nisms such as Probabilistic U-Net [66], hierarchical probabilistic inpainting [109], and conditional flow matching or rectified flows [76, 81, 145] provide principled ways to sample diverse outputs consistent with the observation. Here, by com- puting an observed BEV floormap, the binary BEV floormap completion task is effectively anoutpaintingproblem onbinary metric grid mapsrather than RGB texture synthesis. BEV scene understanding by posterior sampling.Autonomous driving BEV models [73, 103] couple geometric lifting with semantic prediction in synchro- nized multi-camera settings. Geometry-aware multi-view attention further im- proves transformer designs [91]. Neural map priors [155] and generative map layout estimation [165] extend this by incorporating learned spatial priors. Map- grounded visual localization methods such as OrienterNet [120] further show that operational image-to-map alignment is feasible at scale. Indoor BEV work is comparatively sparse: under a single egocentric view, occlusion and limited field of view leave large floor regions unobserved, making completion a partial- observability problem [2, 5, 19, 68]. Closest in spirit to our work, DiffBEV [166] and MapPrior [165] apply conditional diffusion to BEV completion and map layout estimation, but focus on outdoor driving with multi-view camera rigs. In contrast, we instantiate the inpainting and posterior-sampling paradigms for indoorBEV floor completion, where the target is abinarygeometric map that respects metric floor structureâwithout expending capacity on RGB textures or dense semantic labels. A partial BEV observation typically admits multiple plausible completions [4, 44,51,66]. The goal is to sample fromp(F â |F obs ,U)rather than collapse to a point estimate [133,138]. Deterministic predictors target a single mode, whereas stochastic generators represent multimodal uncertainty that is relevant to safety- critical navigation [13,35,68]. Inference efficiency.A large body of work targets dense 3D reconstruction [122, 129,137], including diffusion-based inverse solvers, image and text to mesh and single image novel view synthesis models [22, 54, 80, 82, 93, 113, 119, 130, 140]. These methods are complementary: they optimize 3D geometry or appearance for reconstruction and generation, typically producing volumetric occupancy grids, meshes, or radiance fields with non-trivial inference costs. In contrast, many navigation pipelines use a lightweight2Doccupancy or traversability grid aligned with motion feasibility (typically 2D), with lower storage requirements than Generative Single-View Floormap Completion5 volumetric maps [132,142], motivating our use of BEV representation in indoor scenes. 3 Problem and Framework Afloormapis a binary birdâs-eye-view (BEV) map encoding local traversabil- ity within a bounded indoor region. Binary traversability is the minimal spatial primitive consumed by collision-checking and path-planning modules [41, 85]; richer semantic labels are complementary but orthogonal to the completion task studied here. Given a single egocentric RGB image (Fig. 1), the goal is to pre- dict the complete floormapâincluding unobserved portions that lie outside the camera frustum. We decouple this into two stages:perceptionof the observed floormap from the image, followed bycompletionof the unobserved portion. 3.1 Task: Unobserved Floormap Completion LetRdenote a 2D bounded region of interest, situated relative to an oriented camera pose. Without loss of generality,Ris discretized as anHĂWgrid at a fixed metric resolutionâ(m/px). Given an RGB imageI, the cameraâs field-of- view partitionsRinto anobservedsub-regionR O âRand its complement, the unobservedregionR U =R O . The observed floormapF obs is a binary map overR O that records per-cell traversability,F obs (x) =1(xis traversable)for eachxâR O . In practice,F obs is estimated fromIvia the perception front-end described in Sec. 3.2. The unobserved floormapF uno , defined analogously overR U . Since a single partial observation does not generally determineF uno uniquely, we cast the task as sampling from the posteriorP(F uno |I)as a conditional inverse problem [133]. A parametric networkq Ξ , trained to approximate this posterior, producesK completions Ë F (k) uno âŒq Ξ (I), k= 1,...,K.(1) When the model is deterministic,q Ξ collapses to a point estimate (K=1). 3.2 Estimating the Observed Floormap from an Egocentric View As shown in Fig. 1, the observed BEV floormap is produced from a single RGB image by two deterministic operators [12]. A preprocessing stageΊextracts an image-plane depth mapDvia DepthPro [17] and a floor segmentationSvia SegFormer [153] (trained on ADE-20K [163]). A projection stageΚthen back- projects these signals into the BEV grid using camera intrinsicsK(metric depth is used directly when available), yielding the observed floormapF obs and the observation footprint maskO: (D,S) :=Ί(I),(F obs ,O) :=Κ(D,S;K).(2) HereOandUare the binary masks indicating membership inR O andR U , respectively. At inference, neither the camera pose nor the complete map is 6Bhattacharjee et al. Table 1: Source datasets aggregated into FlatLands.Obs. (raw) are upstream counts before quality filtering; Obs. (filtered) retain only observations with conditional signal ratior cond â„0.10. ScanNet++ is OOD test-only. Details can be found in Sec. S2. DatasetScenes Obs. (raw) Obs. (filtered) Asset Used ZInD [34]7,026168,624133,096 Panorama (see Sec. S2.1) ARKitScenes [7]4,803115,27140,282 PLY mesh Matterport3D [24] 2,10150,42438,004 PLY mesh ScanNet [36]1,50836,19224,763 PLY mesh 3RScan [147]1,29130,98418,216 OBJ mesh ScanNet++ [160]92722,24816,214 PLY mesh (OOD-only) FlatLands(Ours) 17,656 423,743270,5754Ăbinary 2D maps available; completion models operate solely on the abstract BEV conditioning pair(F obs ,U). 3.3 Training Floormap Completion Models Stochastic models drawK>1samples to represent posterior uncertainty. Each sample Ë F (k) uno is assembled into a full completion by preserving observed evidence exactly: F (k) comp =F obs +Uâ Ë F (k) uno .(3) This evidence-clamping step enforces a posterior support constraint: returned completions agree withF obs on all observed cells while remaining free to vary overU. Models are supervised against the ground-truth floormapF â , derived from scene meshes provided by the source datasets. The dataset also provides a valid-workspace maskVthat distinguishes in-bounds cells from out-of-bounds or undefined regions. All losses and metrics are computed only on the unobserved valid evaluation regionR eval =UâV, so scores reflect completion quality beyond the camera frustum. The canonical camera convention and deployed RGB-to-floormaps front-end are specified in Sec. S1. 4 FlatLands Dataset Sources, scope, and canonical splits.FlatLands combines six real indoor metric datasets: Matterport3D [24], ScanNet [36], ScanNet++ [160], ARKitScenes [7], 3RScan [147], and ZInD [34], covering17,656unique scene layouts. We synthesize egocentric observations by sampling floor-valid camera centers and24yaw head- ings per center, with visibility computed via field-of-view and ray-occlusion tests. From423,743synthesized observations, filtering yields a canonical set of270,575 observations (215,342 train, 26,890 val, 28,343 test; Fig. 2a), with ScanNet++ held outentirelyas test-only OOD data. All splits are scene-disjoint. Source aggregation is summarized in Tab. 1 with source-wise breakdown in Fig. 2a. Generative Single-View Floormap Completion7 ZInD 49.2 % ARKit 14.9 % MP3D 14.0 % ScanNet 9.2 % 3RScan 6.7 % ScanNet++ 6.0 % (a)Source dataset breakdown across the six indoor RGBD corpora aggregated in the Flat- Lands dataset. (b)Egocentric BEV from a single camera observa- tion. Visible floor points are orthographically pro- jected onto the floor plane and rasterized on the grid. Fig. 2:FlatLands dataset statistics and construction. RGB inputF obs UF â V Fig. 3:Input egocentric RGB (left) and the four aligned256Ă256binary maps per observation.F obs : observed floor;U: valid unobserved;F â : full floor ground truth;V: valid workspace. The white marker (âŒ) denotes the fixed camera anchor in BEV. Data construction pipeline.Each observation is generated offline by placing a vir- tual camera in the reconstructed 3D mesh at the sampled pose, back-projecting the rendered view, and orthogonally projecting the visible floor into a256Ă256 egocentric BEV grid as shown in Fig. 2b. The agentâs understanding of its pose in the world is not assumed at inference (egocentric). Full processing details, in- cluding observation synthesis and crop validation, are in Secs. S2, S2.2 and S2.4. The pipeline outputs four aligned binary maps per observation (Fig. 3):F obs (observed floor),U(valid but unobserved),F â (full floor ground truth), andV (valid workspace). By construction,F obs âȘUâV. Depth and mesh assets are used only offline: floor evidence comes from mesh plane-fitting for five datasets and from DA 2 metric depth [70] for ZInD panoramas, calibrated via room-vertex annotations. Camera intrinsics, anchor convention, and grid resolution are fixed, enabling consistent difficulty interpretation across sources and ID/OOD subsets. Source-specific processing, aggregation and filtering rules are in Secs. S2.1 and S2.4; split consistency and scene-level statistics are in Sec. S2.5. OnR eval , floor prevalence is moderately high (meanâ0.80); its effect on metric ranking is analyzed in Sec. S2.4. Canonical tensors use a256Ă256grid at 25.6 px/m, i.e., ââ0.039m per pixel; the cell-size rationale is explained in Secs. S2.3 and S2.4. 8Bhattacharjee et al. 5 Experiments We benchmark11methods spanning parameter-free approaches, deterministic completion, an epistemic ensemble, and stochastic posterior samplers. To isolate modeling effects, all learned methods share the same interface and evaluation protocol: inputs are(F obs ,U); losses and metrics are computed on the common supervision maskR eval . A compact per-method summary is provided in Table 2. 5.1 Baselines Naive baselines.Four parameter-free methods provide reference fill strategies forU, then are merged with observed evidence using the hard-clamp rule [86] (Eq. (3)).All Obstaclepredicts obstacle (0) everywhere onU.All Floor predicts floor (1) everywhere onU.N Propagationassigns an unobserved cell the label of the nearest observed one.Uniform Randomsamples i.i.d. Bernoulli(0.5)labels per cell. Deterministic models.We evaluateUNet[117,151],PartialConv UNet[79], andLaMa[31,135]. All three keep their original backbones; we only adapt the in- put and output channels to binary BEV floor completion. UNet and PConv UNet are trained with masked BCE reconstruction losses onR eval , while LaMa retains its reconstruction-plus-adversarial objective. At inference, continuous outputs are binarized via fixed thresholding [27] and then hard-clamped to preserve ob- served evidence [86]. Stochastic models.LaMa Ensembletrains four independent LaMa models with different random seeds [4, 51] and returns one output from each mem- ber (K=4), providing seed-based diversity. Increasing the number of ensemble samples requires retraining a model on a different seed. We also evaluate three posterior conditional samplers:Diffusion[52,127],Flow Matching[76,77,81], andFlow Matching with Cross-Attention Conditioning (FM+XAttn). FM+XAttn keeps the baseline concatenated input and adds a condition en- coder on[F obs ,U], with cross-attention [146] injected only at coarse resolutions (64Ă64 and 32Ă32), following modular conditioning ideas [102,116,162]. Operat- ing cross-attention at coarse resolutions keeps parameter overhead modest while letting the generator attend to global layout cues from the conditioning pair; fine-grained spatial detail is resolved by the convolutional decoder. All three samplers share the same setup: masked BCE onR eval , classifier-free guidance (s=2.0) [37, 53], per-step evidence clamping [32, 115],K=4samples, and itera- tive sampling (50DDIM steps for Diffusion,50Heun steps for Flow Matching, 25Heun steps for FM+XAttn [50,62,127]). Implementation and reproducibility.All models are trained in distributed full- precision mode on4ĂA100 40 GB GPUs under identical schedules; inference runs on consumer-grade hardware (Nvidia RTX 4090). Further architecture, training, and sampling details are reported in Secs. S5 and S8.2 and Tab. S12. Generative Single-View Floormap Completion9 Table 2: Baseline catalog.Unless noted, learned methods share inputs(F obs ,U), supervision maskR eval , and hard evidence clamping before evaluation. MethodAdaptationsInference All naive baselinesAll-Obstacle (fill 0), All-Floor (fill 1), N Propagation (nearest-observed copy), Uniform Random (p=0.5). Parameter-free rules onU; merged with observed evidence via hard clamping (Eq. (3)). Single pass U-Net [117,151]Masked BCE onR eval . Backbone unchanged; binary BEV I/O only.Thr. 0.5 + clamp PConv-UNet [79]Masked BCE onR eval . Partial-conv. mask propagation retained; adapted to binary BEV completion. Thr. 0.5 + clamp LaMa [31,135]Recon. + adversarial. Fourier-conv. generator retained; channels remapped for floor-occupancy completion. Thr. 0.5 + clamp LaMa Ens. [4,51]Same as LaMa (4 seeds). Epistemic variant from independent seed train- ing; no architecture change. K=4 Diffusion [52,127] Masked BCE onR eval . Posterior inpainting with per-step evidence clamping [32,115]. DDIM, 50 st., CFGs=2 Flow Match. [76]Masked BCE onR eval . Same conditioning/clamping as Diffusion; ODE transport parameterization. Heun, 50 st., CFGs=2 FM+XAttn [76,146] Masked BCE onR eval . Adds condition encoder on[F obs ,U]with cross- attn. (64Ă64,32Ă32); see [102,116,162]. Heun, 25 st., CFGs=2 The source code, pre-trained weights, and the full dataset reproduction pipeline will be publicly released. 5.2 Evaluation Protocol We evaluate on the canonical test split (N=28,343), partitioned into12,129in- distribution observations and16,214out-of-distribution observations. Quantita- tive results are computed on standardized BEV-conditioned inputs. Difficulty- tier and threshold analyses are in Sec. S2.5. All metrics are computed on the unobserved valid regionR eval =UâV. We additionally report the standard metrics IoU and F1 with traversable floor as the positive class (see Sec. S6). For this benchmark, we introduce two new metricsâUMR and MESâdesigned to calibrate fidelity and diversity under masked evaluation. Unobserved-region mismatch rate.We define the unobserved-region mismatch rate (UMR; lower is better) as the fraction of misclassified cells onR eval : UMR = FP + FN |R eval | = 1â TP + TN |R eval | .(4) Since floor prevalence onR eval is high (â0.80; a natural bias in indoor spaces; see Sec. S2.4 for details), IoU and F1 primarily reflect floor-completion fidelity and can remain high even when wall boundaries are imperfect. Masked Energy Score.For multi-sample methods, we introduce the Masked En- ergy Score (MES; lower is better) [46,114]: MES = 1 K K X k=1 d R eval F (k) comp ,F â â 1 2K(Kâ1) X kÌž=â d R eval F (k) comp ,F (â) comp , (5) 10Bhattacharjee et al. Table 3:Fidelity metrics (mean±std) onR eval with oracle best-of-K(K=4) variant for stochastic methods (bottom four rows). MethodIDOOD UMRâIoUâF1âUMRâIoUâF1â All-Obstacle0.661±0.229 0.000±0.000 0.000±0.000 0.854±0.155 0.000±0.000 0.000±0.000 All-Floor0.339±0.229 0.661±0.229 0.737±0.210 0.146±0.155 0.854±0.155 0.912±0.116 N Prop.0.243±0.078 0.620±0.219 0.708±0.202 0.198±0.156 0.779±0.186 0.861±0.143 Uniform Rand.0.501±0.002 0.367±0.104 0.516±0.128 0.502±0.003 0.453±0.057 0.621±0.064 U-Net0.122±0.030 0.763±0.132 0.843±0.098 0.114±0.111 0.870±0.136 0.924±0.094 PConv-UNet0.126±0.033 0.758±0.132 0.838±0.098 0.114±0.113 0.870±0.139 0.924±0.097 LaMa0.209±0.059 0.623±0.161 0.734±0.131 0.283±0.159 0.672±0.190 0.787±0.153 LaMa-Ens.0.105±0.099 0.817±0.205 0.882±0.154 0.107±0.094 0.874±0.126 0.927±0.083 Diffusion0.099±0.097 0.830±0.196 0.892±0.143 0.094±0.092 0.890±0.122 0.936±0.082 Flow Match.0.100±0.095 0.829±0.1920.892±0.1390.102±0.099 0.880±0.129 0.930±0.088 FM+XAttn0.100±0.0940.830±0.192 0.892±0.1390.109±0.092 0.873±0.124 0.927±0.084 Table 4:Stochastic evaluation (mean±std) onR eval forK=4. MES: Masked Energy Score; IoU m : mean-of-KIoU; Var: average per-pixel variance. MethodIDOOD MESâIoU m âVarâMESâIoU m âVarâ LaMa-Ens.0.158±0.139 0.713±0.246 0.076±0.062 0.149±0.093 0.720±0.155 0.084±0.043 Diffusion0.143±0.152 0.763±0.222 0.044±0.029 0.097±0.1020.832±0.1240.048±0.026 Flow Match.0.143±0.1480.766±0.226 0.037±0.0290.101±0.1030.832±0.135 0.043±0.028 FM+XAttn0.133±0.138 0.766±0.2180.044±0.0320.095±0.0930.813±0.123 0.061±0.028 withd R eval (A,B) = 1âIoU R eval (A,B)(Jaccard distance [75]), which normal- izes by the union and shares the same geometric semantics as the fidelity axis (see Sec. S6). The first term penalizes distance to the ground truth (fidelity); the second rewards pairwise spread among samples (diversity). We setK= 4 (Sec. S7.2); sample-count sensitivity is analyzed in Sec. S7.2. Main tables report mean±std. We also report mean IoU, best-of-KIoU, and per-pixel variance for the stochastic methods. For fairness, all methods use identical post-processing: fixed map thresholding [27], evidence hard clamping, and the common evalua- tion maskR eval , preventing protocol differences from affecting model ranking. Downstream integration into closed-loop planners is an important direction but lies beyond the scope of this benchmark; we focus on fidelity and calibration of the completion stage itself. 6 Results and Discussion Table 3 reports UMR, IoU, and F1 (mean±std) onR eval . Among deterministic predictors on the ID split, UNet and PConv UNet lead, while LaMa trailsâ consistent with adversarial training trading distortion for perceptual realism [16]. The trivial All Floor baseline reaches IoU comparable to LaMa because high floor prevalence (âŒ0.80; Sec. S2.4) inflates IoU for optimistic guesses (Sec. 5.2); N Propagation ranks lower because nearest-neighbor copying replicates boundary pixels, lowering recall. Generative Single-View Floormap Completion11 LaMa-Ensemble FM+XAttn samples F obsU F â V Ï 2 Ï 2 Fig. 4:LaMa-Ensemble vs. FM+XAttn on a multi-room ScanNet scene.Row 1: ob- served floorF obs and unobserved maskUcondition both models; the four LaMa- Ensemble samples (boxed) and their per-pixel varianceÏ 2 .Row 2: ground-truth floor F â and validity maskVused for evaluation; four FM+XAttn samples (boxed) and theirÏ 2 . LaMa-Ensemble spreads variance uniformly; FM+XAttn concentrates it at layout boundaries. With oracle best-of-Kselection (K=4), all stochastic generators surpass ev- ery deterministic predictor, confirming that posterior sampling recovers higher- fidelity completions. Method ordering is stable despite per-scene variance, with stochastic best-of-Kreducing UMR furthest. Oracle selection is a standard di- agnostic for posterior coverage [46, 66]; the oracle-free MES in Tab. 4 confirms the same ranking without privileged access to the ground truth. On the OOD split (ScanNet++), all methods improve in IoU and the ID ranking is pre- served (Sec. S7), validating cross-source generalization. Since ScanNet++ con- tains room geometries and capture conditions unseen during training, the pre- served ranking indicates that learned layout priors transfer across architectural styles rather than overfitting to source-specific artifacts. Additional quantita- tive analysis is reported in Sec. S4. Table 4 shows FM+XAttn achieves the best MES on both splits, with Flow Matching and Diffusion close behind and LaMa-Ensemble trailing. Mean-of-KIoU is nearly identical across continuous generators, so differences lie in calibration rather than posterior mean accuracy; FM+XAttn trades slightly higher variance for improved calibration. Boundary variance.Following Chenget al. [29], we partition unobserved floor pixels into an interior set⊠int (â„7px from GT non-floor) and a boundary set ⊠bnd (within7px of a floorânon-floor transition; justification in Sec. S3). A cali- brated sampler should produce near-zero variance on⊠int ; boundary variance re- flects genuine layout ambiguity. FM+XAttn attains ÌÏ 2 int = 6Ă10 â5 , ÌÏ 2 bnd = 0.037 (ratioâŒ600Ă), concentrating uncertainty at boundaries. LaMa-Ensemble yields ÌÏ 2 int =0.052and ÌÏ 2 bnd =0.125(ratio2.4Ă): its interior variance isâŒ880Ălarger than that of FM+XAttn, indicating seed-level divergence rather than poste- rior ambiguity (see Fig. 4). FM+XAttn thus localizes uncertainty to genuinely ambiguous boundaries, whereas LaMa-Ensemble inflates variance in geometri- 12Bhattacharjee et al. F obs UGTAll-Obstacle All-Floor N-Prop UniformU-NetPConvLaMa LaMa-EnsDiffusion Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 FlowFM+XAttn Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Fig. 5: Qualitative results on the test split. Top:deterministic single-output comparison across three scenes (in-distribution rows 1â2, out-of-distribution row 3). Columns show the observed floorF obs , unobserved maskU, ground truth, and pre- dictions from each baseline. These BEV observations are geometrically projected from the 3D mesh and do not involve any RGB input.Bottom:four independent samples, drawn from each stochastic generator for one in-distribution scene, alongside the per- pixel varianceÏ 2 (brighter=higher disagreement). cally determined interiors, raising MES. For a downstream planner, boundary- concentrated uncertainty is directly actionable: high-variance cells flag regions where collision risk is ambiguous and information-gathering actions would reduce planning uncertainty most. Qualitative observations.Figure 5 shows representative test scenes: determin- istic models are sharp in open regions but hallucinate non-floors at structural ambiguities, whereas stochastic generators spread mass across plausible layouts and highlight decision-relevant uncertainty (extended in Sec. S7). Fig. 6 evaluates the monocular RGBâfloormaps pipeline on held-out Scan- Net++ images using Depth Pro [17] and SegFormer [153]. All learned models remain coherent under estimated BEV conditioning, confirming transfer from ground-truth inputs. The front-end projection introduces partial-depth artifacts and frustum-edge gaps, yet stochastic generators still produce plausible room ex- tensions where deterministic methods truncate or fragment the floor. Comparing variance maps between Fig. 5 and Fig. 6, front-end noise slightly inflates inte- rior variance but boundary-concentrated structure is preserved, indicating layout priors robust to moderate input corruption. Additional quantitative results, in- Generative Single-View Floormap Completion13 RGBF obs UGT All-Obstacle All-Floor N-Prop UniformU-NetPConvLaMa LaMa-EnsDiffusion Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 FlowFM+XAttn Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Ë F (1) Ë F (2) Ë F (3) Ë F (4) Ï 2 Fig. 6: End-to-end inference pipeline results (RGB input). Top:determinis- tic single-output comparison across three ScanNet++ scenes processed through the full monocular RGBâfloormaps pipeline (Fig. 1). The RGB column shows the input egocentric image; the projectedF obs is noisier than the mesh-derived case (Fig. 5), increasing completion difficulty.Bottom:four independent posterior samples Ë F (k) uno , k=1,...,4, drawn from each stochastic generator for one scene, alongside the per-pixel varianceÏ 2 (brighter=higher disagreement). cluding end-to-end quantitative results, a synthetic ambiguity study when the true posterior under evaluation is fixed, and further qualitative examples are in Secs. S4 and S7. These observations hold under favorable conditioning; we next examine how all methods degrade when conditioning becomes scarce. Failure modes.Fig. 7 isolates a hard case where a broad occlusion wedge leaves over half the GT floor outside the observation (red overlay:F â obs ). Because no conditioning signal constrains this region, floor-boundary placement is geomet- rically under-determined, and most methods exhibit boundary leakageâover- extending floor into obstacle-heavy areas. In this specific case LaMa instead col- lapses (IoU=0.003): with sparse conditioning, its adversarial objective becomes unstableâthe discriminator trivially rejects large-region completions, causing the generator to mode-collapse to near-zero output [47]. Two additional fail- ure patterns appear in the extended grid (Sec. S7.3): disconnected floor islands from sparse conditioning, and residual boundary artifacts in otherwise high-IoU scenes. 14Bhattacharjee et al. Predictions F obsF â F â obs DiffusionFlowFM+XAttnLaMa Fig. 7: Failure case: boundary leakage under wide occlusion.The red over- lay (F â obs ) highlights GT floor absent from the observationâover half the true floor hereâleaving boundary placement geometrically under-determined. All genera- tive methods over-extend floor into this ambiguous region; FM+XAttn is most con- servative but residual leakage persists. LaMa collapses entirely (IoU=0.003); one such other example of collapse is in Fig. 6 row 3. Extended grid in Sec. S7.3. 7 Conclusion: Limitations & Future Work We presented a unified benchmark for single-view BEV floor completion with identical conditioning, masking, and scoring across deterministic, ensemble, and stochastic methods, with explicit ID and OOD splits. Three findings emerge: (1) stochastic generators with oracle selection surpass all deterministic predic- tors, confirming that posterior sampling recovers completions no single-pass estimator can match; (2) FM+XAttn concentrates variance at layout bound- aries while ensemble seeds diverge globallyâepistemic spread is a poor proxy for aleatoric layout ambiguity; (3) method ordering is preserved on the OOD split, indicating stable generalization. This benchmark targetsbinarytraversability only, without semantic cues [8, 129,166], and assumes simplified pinhole intrinsics that do not capture the diver- sity of real robotic cameras. It also omitsgradedtraversability (e.g. stairs, slopes, etc.), which would better reflect the feasible traversability constraints of hetero- geneous embodied platforms like mobile, legged, or aerial robots. Failure analysis (Fig. 7; Sec. S7.3) indicates that when most ground-truth floor is unobserved, boundary placement becomes under-determined; consequently, all methods pro- duce local defects poorly captured by IoU alone. Promising directions include expanding the source datasets to improve architectural diversity and building ty- pologies; extending to multi-storey and multi-room settings; incorporating RGB and semantic cues for conditioning; adopting traversability definitions for differ- ent robotic systems; exploring modern backbones [28,39,101,146,154,157]; and moving beyond single-view input to multi-view settings. Ethical considerations are discussed in Sec. S9. Completed floormaps can serve as probabilistic spatial priors for belief-space planning [12,35,48,68], active perception [45,110], SLAM back-ends [19,40,89, 120], and generative motion planning [20,85,134,137]. Because the representation is an abstract binary grid, the completion stage is sensor-agnostic. Closed-loop evaluation under map uncertainty [35,104,121] is a natural next step extending to planning and navigation [4,14,51]. Generative Single-View Floormap Completion15 References 1. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R.: Structured Denois- ing Diffusion Models in Discrete State-Spaces. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) 2. Axelrod, B., Kaelbling, L.P., Lozano-PĂ©rez, T.: Provably Safe Robot Navigation with Obstacle Uncertainty. The International Journal of Robotics Research (2018) 3. Aydemir, A., Jensfelt, P., Folkesson, J.: What Can We Learn from 38,000 Rooms? Reasoning about Unexplored Space in Indoor Environments. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2012) 4. Baek, S., Moon, B., Kim, S., Cao, M., Ho, C., Scherer, S., Jeon, J.: PIPE Planner: Pathwise information gain with map predictions for indoor robot exploration. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2025) 5. Banfi, J., Woo, L., Campbell, M.: Is it Worth to Reason about Uncertainty in Occupancy Grid Maps during Path Planning? In: Proceedings of the IEEE Inter- national Conference on Robotics and Automation (ICRA) (2022) 6. Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric Corre- spondence and Chamfer Matching: Two New Techniques for Image Matching. In: International Joint Conference on Artificial Intelligence (IJCAI) (1977) 7. Baruch, G., Chen, Z., Dehghan, A., Dimry, T., Feigin, Y., Fu, P., Gebauer, T., Kurz, D., Joffe, B., Schwartz, A., Shulman, E.: ARKitScenes: A diverse real- world dataset for 3D indoor scene understanding using mobile RGB-D data. In: Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track (2021) 8. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A dataset for semantic scene understanding of LiDAR se- quences. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) (2019) 9. Bengio, Y., LĂ©onard, N., Courville, A.: Estimating or Propagating Gradi- ents Through Stochastic Neurons for Conditional Computation. arXiv preprint arXiv:1308.3432 (2013) 10. Bercher, P., Alford, R., Höller, D.: A Survey on Hierarchical Planning â One Abstract Idea, Many Concrete Realizations. In: Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) (2019) 11. Bertalmio, M., Sapiro, G., Caselles, V., Ballester, C.: Image Inpainting. In: Pro- ceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH) (2000) 12. Bhattacharjee, S.S., Lu, H., Campbell, D., Shome, R.: Into the Unknown: To- wards using generative models for sampling priors of environment uncertainty for planning in configuration spaces. arXiv preprint arXiv:2510.11014 (2025) 13. Bhattacharjee, S.S., Campbell, D., Shome, R.: Believing is Seeing: Unobserved Object Detection using Generative Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 14. Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., Siegwart, R.: Receding Hori- zon "Next-Best-View" Planner for 3D Exploration. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2016) 15. Black Forest Labs, Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, 16Bhattacharjee et al. C., Lorenz, D., MĂŒller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space. arXiv preprint arXiv:2506.15742 (2025) 16. Blau, Y., Michaeli, T.: The Perception-Distortion Tradeoff. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 17. Bochkovskii, A., Delaunoy, A., Germain, H., Santos, M., Zhou, Y., Richter, S.R., Koltun, V.: Depth Pro: Sharp Monocular Metric Depth in Less Than a Second. In: International Conference on Learning Representations (ICLR) (2025) 18. Bunse-Gerstner, A., Byers, R., Mehrmann, V., Nichols, N.K.: Numerical compu- tation of an analytic singular value decomposition of a matrix valued function. Numerische Mathematik (1991) 19. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age. IEEE Transactions on Robotics (2016) 20. Carvalho, J., Le, A.T., Baierl, M., Koert, D., Peters, J.: Motion Planning Diffu- sion: Learning and Planning of Robot Motions with Diffusion Models. In: Pro- ceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2023) 21. Carvalho, J., Le, A.T., Kicki, P., Koert, D., Peters, J.: Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models. IEEE Transactions on Robotics (2025) 22. Chan, E.R., Nagano, K., Chan, M.A., Bergman, A.W., Park, J.J., Levy, A., Ait- tala, M., Mello, S.D., Karras, T., Wetzstein, G.: Generative Novel View Synthesis with 3D-Aware Diffusion Models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 23. Chan, T.F., Shen, J.: Nontexture Inpainting by Curvature-Driven Diffusions. Journal of Visual Communication and Image Representation (2001) 24. Chang, A.X., Dai, A., Funkhouser, T., Halber, M., NieĂner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from RGB-D data in indoor environments. In: International Conference on 3D Vision (3DV) (2017) 25. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-SP: Inverse CAD for Floorplans by Sequential Room-wise Shortest Path. In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) (2019) 26. Chen, T., Xu, B., Zhang, C., Guestrin, C.: Training Deep Nets with Sublinear Memory Cost. arXiv preprint arXiv:1604.06174 (2016) 27. Chen, T., Zhang, R., Hinton, G.: Analog bits: Generating discrete data using diffusion models with self-conditioning. In: International Conference on Learning Representations (ICLR) (2023) 28. Chen, X., Mishra, N., Rohaninejad, M., Abbeel, P.: PixelSNAIL: An improved autoregressive generative model. In: Proceedings of the International Conference on Machine Learning (ICML) (2018) 29. Cheng, B., Girshick, R., DollĂĄr, P., Berg, A.C., Kirillov, A.: Boundary IoU: Im- proving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 30. Cheng, Y.C., Lin, C.H., Lee, H.Y., Ren, J., Tulyakov, S., Yang, M.H.: In&out: Diverse image outpainting via GAN inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) Generative Single-View Floormap Completion17 31. Chi, L., Jiang, B., Mu, Y.: Fast Fourier Convolution. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) 32. Christopher, J.K., Baek, S., Fioretto, F.: Constrained Synthesis with Pro- jected Diffusion Models. In: Advances in Neural Information Processing Systems (NeurIPS) (2024) 33. Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion Posterior Sampling for General Noisy Inverse Problems. In: International Conference on Learning Representations (ICLR) (2023) 34. Cruz, S., Hutchcroft, W., Li, Y., Khosravan, N., Boyadzhiev, I., Kang, S.B.: Zillow Indoor Dataset: Annotated floor plans with 360 ⊠panoramas and 3D room layouts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 35. Curtis, A., Matheos, G., Gothoskar, N., Mansinghka, V., Tenenbaum, J.B., Lozano-PĂ©rez, T., Kaelbling, L.P.: Partially Observable Task and Motion Plan- ning with Uncertainty and Risk Awareness. In: Robotics: Science and Systems (RSS) (2024) 36. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., NieĂner, M.: Scan- Net: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 37. Dhariwal, P., Nichol, A.: Diffusion Models Beat GANs on Image Synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) 38. Dice, L.R.: Measures of the amount of ecologic association between species. Ecol- ogy (1945) 39. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In: International Conference on Learning Representations (ICLR) (2021) 40. Durrant-Whyte, H., Bailey, T.: Simultaneous Localization and Mapping: Part I. IEEE Robotics & Automation Magazine (2006) 41. Elfes, A.: Using Occupancy Grids for Mobile Robot Perception and Navigation. Computer (1989) 42. Fang, X., Garrett, C.R., Eppner, C., Lozano-PĂ©rez, T., Kaelbling, L.P., Fox, D.: DiMSam: Diffusion Models as Samplers for Task and Motion Planning under Partial Observability. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2024) 43. Fischler, M.A., Bolles, R.C.: Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography. Com- munications of the ACM (1981) 44. Gal, Y., Ghahramani, Z.: Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. In: Proceedings of the International Con- ference on Machine Learning (ICML) (2016) 45. Garg, S., SĂŒnderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., Corke, P., Milford, M.: Semantics for Robotic Mapping, Perception and Interaction: A Survey. Foundations and TrendsÂźin Robotics (2020) 46. Gneiting, T., Raftery, A.E.: Strictly Proper Scoring Rules, Prediction, and Esti- mation. Journal of the American Statistical Association (2007) 47. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative Adversarial Nets. In: Advances in Neural Information Processing Systems (NeurIPS) (2014) 18Bhattacharjee et al. 48. Han, Y., Banfi, J., Campbell, M.: Planning Paths Through Unknown Space by Imagining What Lies Therein. In: Conference on Robot Learning (CoRL) (2020) 49. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cam- bridge University Press (2004) 50. Heun, K.: Neue methode zur approximativen lösung der differentialgleichungen einer unabhĂ€ngigen verĂ€nderlichen. Zeitschrift fĂŒr Mathematik und Physik (1900) 51. Ho, C., Kim, S., Moon, B., Parandekar, A., Harutyunyan, N., Wang, C., Sycara, K., Best, G., Scherer, S.: MapEx: Indoor structure exploration with probabilis- tic information gain from global map predictions. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2025) 52. Ho, J., Jain, A., Abbeel, P.: Denoising Diffusion Probabilistic Models. In: Ad- vances in Neural Information Processing Systems (NeurIPS) (2020) 53. Ho, J., Salimans, T.: Classifier-Free Diffusion Guidance. In: NeurIPS Workshop on Diffusion Models (2022) 54. Hollein, L., Cao, A., Owens, A., Johnson, J., Niessner, M.: Text2Room: Extract- ing Textured 3D Meshes from 2D Text-to-Image Models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 55. Hoogeboom, E., Heek, J., Salimans, T.: Simple Diffusion: End-to-End Diffusion for High Resolution Images. In: Proceedings of the International Conference on Machine Learning (ICML) (2023) 56. Hoogeboom, E., Nielsen, D., Jaini, P., ForrĂ©, P., Welling, M.: Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) 57. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., Burgard, W.: OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous Robots (2013) 58. Hua, B.S., Pham, Q.H., Nguyen, D.T., Tran, M.K., Yu, L.F., Yeung, S.K.: Sce- neNN: A scene meshes dataset with annotations. In: International Conference on 3D Vision (3DV) (2016) 59. Jaccard, P.: Ătude comparative de la distribution florale dans une portion des Alpes et des Jura. Bulletin de la SociĂ©tĂ© Vaudoise des Sciences Naturelles (1901) 60. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual Losses for Real-Time Style Transfer and Super-Resolution. In: Proceedings of the European Conference on Computer Vision (ECCV) (2016) 61. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and Acting in Partially Observable Stochastic Domains. Artificial Intelligence (1998) 62. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the Design Space of Diffusion-Based Generative Models. In: Advances in Neural Information Process- ing Systems (NeurIPS) (2022) 63. Katsumata, Y., Kanechika, A., Taniguchi, A., Hafi, L.E., Hagiwara, Y., Taniguchi, T.: Map Completion from Partial Observation using the Global Structure of Mul- tiple Environmental Maps. Advanced Robotics (2022) 64. Katyal, K., Popek, K., Paxton, C., Burlina, P., Hager, G.D.: Uncertainty-Aware Occupancy Map Prediction Using Generative Networks for Robot Navigation. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2019) 65. Kerbl, B., Kopanas, G., LeimkĂŒhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (2023) 66. Kohl, S.A.A., Romera-Paredes, B., Meyer, C., Fauw, J.D., Ledsam, J.R., Maier- Hein, K.H., Eslami, S.M.A., Rezende, D.J., Ronneberger, O.: A Probabilistic U- Generative Single-View Floormap Completion19 Net for Segmentation of Ambiguous Images. In: Advances in Neural Information Processing Systems (NeurIPS) (2018) 67. Kuipers, B.: The Spatial Semantic Hierarchy. Artificial Intelligence (2000) 68. Kurniawati, H.: Partially Observable Markov Decision Processes and Robotics. Annual Review of Control, Robotics, and Autonomous Systems (2022) 69. LaValle, S.M.: Planning Algorithms. Cambridge University Press (2006) 70. Li, H., Zheng, W., He, J., Liu, Y., Lin, X., Yang, X., Chen, Y.C., Guo, C.: DA 2 : Depth anything in any direction. In: International Conference on Learning Rep- resentations (ICLR) (2026) 71. Li, J., Chen, C., Xiong, Z.: Contextual outpainting with object-level contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 72. Li, X., Qiu, H., Wang, L., Zhang, H., Qi, C., Han, L., Xiong, H., Li, H.: Challenges and Trends in Egocentric Vision: A Survey. Machine Intelligence Research (2026) 73. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: BEV- Former: Learning Birdâs-Eye-View representation from multi-camera images via spatiotemporal transformers. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 74. Ling, L., Sheng, Y., Tu, Z., Zhao, W., Xin, C., Wan, K., Yu, L., Guo, Q., Yu, Z., Lu, Y., Li, X., Sun, X., Ashok, R., Mukherjee, A., Kang, H., Kong, X., Hua, G., Zhang, T., Benes, B., Bera, A.: DL3DV-10K: A large-scale scene dataset for deep learning-based 3D vision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 75. Lipkus, A.H.: A proof of the triangle inequality for the Tanimoto distance. Journal of Mathematical Chemistry (1999) 76. Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow Matching for Generative Modeling. In: International Conference on Learning Representations (ICLR) (2023) 77. Lipman, Y., Havasi, M., Holderrieth, P., Shaul, N., Le, M., Karrer, B., Chen, R.T.Q., Lopez-Paz, D., Ben-Hamu, H., Gat, I.: Flow Matching Guide and Code. arXiv preprint arXiv:2412.06264 (2024) 78. Liu, C., Wu, J., Furukawa, Y.: FloorNet: A unified framework for floorplan recon- struction from 3D scans. In: Proceedings of the European Conference on Com- puter Vision (ECCV) (2018) 79. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image In- painting for Irregular Holes Using Partial Convolutions. In: Proceedings of the European Conference on Computer Vision (ECCV) (2018) 80. Liu, M., Xu, C., Jin, H., Gao, L., Tao, X., Li, Y.: One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 81. Liu, X., Gong, C., Liu, Q.: Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In: International Conference on Learning Rep- resentations (ICLR) (2023) 82. Liu, Y., Xu, K., Liu, Y., Liu, Y., He, J., Tong, R.: Acc3D: Accelerating Single Image to 3D Diffusion Models via Edge Consistency Guided Score Distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 83. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: International Conference on Learning Representations (ICLR) (2017) 20Bhattacharjee et al. 84. Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: Interna- tional Conference on Learning Representations (ICLR) (2019) 85. Lu, H., Kurniawati, H., Shome, R.: Sampling-Based Motion Planning for Optimal Probability of Collision under Environment Uncertainty. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2024) 86. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: RePaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 87. Luo, Z., Huang, W.: FloorPlanGAN: Vector residential floorplan adversarial gen- eration. Automation in Construction (2022) 88. Maciejewski, A.A., Fox, J.J.: Path planning and the topology of configuration space. IEEE Transactions on Robotics and Automation (1993) 89. Maggio, D., Lim, H., Carlone, L.: VGGT-SLAM: Dense RGB SLAM optimized on the SL(4) manifold. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 90. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision (ECCV) (2020) 91. Miyato, T., Jaeger, B., Welling, M., Geiger, A.: GTA: A geometry-aware attention mechanism for multi-view transformers. In: International Conference on Learning Representations (ICLR) (2024) 92. MĂŒller, M., Casser, V., Lahoud, J., Smith, N., Ghanem, B.: Learning to Find Good Correspondences for Image-Based Room Layout Estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 93. MĂŒller, N., Schwarz, K., Rössle, B., Porzi, L., BulĂČ, S.R., NieĂner, M., Kontschieder, P.: MultiDiff: Consistent novel view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 94. Mumford, D.: Pattern Theory: A Unifying Perspective. In: First European Congress of Mathematics, Paris, July 6â10, 1992, Vol. I. BirkhĂ€user Basel (1994) 95. Nichol, A.Q., Dhariwal, P.: Improved Denoising Diffusion Probabilistic Models. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 96. Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.: Voxblox: Incremental 3D euclidean signed distance fields for on-board MAV planning. In: Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2017) 97. OâMeadhra, C., Tabib, W., Michael, N.: Variable Resolution Occupancy Mapping Using Gaussian Mixture Models. IEEE Robotics and Automation Letters (2019) 98. Papadakis, P.: Terrain Traversability Analysis Methods for Unmanned Ground Vehicles: A Survey. Engineering Applications of Artificial Intelligence (2013) 99. Pathak, D., KrĂ€henbĂŒhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context En- coders: Feature Learning by Inpainting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 100. Paulius, D., Sun, Y.: A Survey of Knowledge Representation in Service Robotics. Robotics and Autonomous Systems (2019) 101. Peebles, W., Xie, S.: Scalable Diffusion Models with Transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) Generative Single-View Floormap Completion21 102. Perez, E., Strub, F., de Vries, H., Dumoulin, V., Courville, A.: FiLM: Visual rea- soning with a general conditioning layer. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2018) 103. Philion, J., Fidler, S.: Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. In: Proceedings of the European Confer- ence on Computer Vision (ECCV) (2020) 104. Platt, R., Tedrake, R., Kaelbling, L.P., Lozano-PĂ©rez, T.: Belief Space Planning Assuming Maximum Likelihood Observations. In: Robotics: Science and Systems (RSS) (2010) 105. Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G.M., Damen, D., Tommasi, T.: An Outlook into the Future of Egocentric Vision. In- ternational Journal of Computer Vision (IJCV) (2024) 106. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., MĂŒller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution im- age synthesis. In: International Conference on Learning Representations (ICLR) (2024) 107. Pronobis, A., Sjö, K., Aydemir, A., Bishop, A.N., Jensfelt, P.: A Framework for Robust Cognitive Spatial Mapping. In: Proceedings of the International Confer- ence on Advanced Robotics (ICAR) (2009) 108. Qi, S., Zhu, Y., Huang, S., Jiang, C., Zhu, S.C.: Human-centric Indoor Scene Syn- thesis Using Stochastic Grammar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 109. Rahman, M., Carneiro, G.: Hierarchical Probabilistic Ultrasound Image Inpaint- ing via Variational Inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 110. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Occupancy Anticipation for Efficient Exploration and Navigation. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020) 111. Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-Matterport 3D Dataset (HM3D): 1000 Large- Scale 3D Environments for Embodied AI. In: Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track (2021) 112. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchi- cal Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125 (2022) 113. Reed, A., Crowe, B., Albin, D., Achey, L., Hayes, B., Heckman, C.: SceneSense: Diffusion models for 3D occupancy synthesis from partial observation. In: Pro- ceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2024) 114. Rizzo, M.L., SzĂ©kely, G.J.: Energy Distance. Wiley Interdisciplinary Reviews: Computational Statistics (2016) 115. Rochman-Sharabi, O., Louppe, G.: Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints. arXiv preprint arXiv:2601.21033 (2026) 116. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-Resolution Image Synthesis with Latent Diffusion Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 117. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2015) 22Bhattacharjee et al. 118. Rottmann, A., Mozos, Ă.M., Stachniss, C., Burgard, W.: Semantic Place Clas- sification of Indoor Environments with Mobile Robots Using Boosting. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2005) 119. Sargent, K., Li, Z., Shah, T., Herrmann, C., Yu, H.X., Zhang, Y., Chan, E.R., Lagun, D., Fei-Fei, L., Sun, D., Wu, J.: ZeroNVS: Zero-shot 360-degree view synthesis from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 120. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: OrienterNet: Visual localization in 2D public maps with neural matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 121. Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A platform for embodied AI research. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV) (2019) 122. Schmid, L., Cheema, M.N., Reijgwart, V., Siegwart, R., Tombari, F., Cadena, C.: SC-Explorer: Incremental 3D scene completion for safe and efficient exploration mapping and planning. arXiv preprint arXiv:2208.08307 (2022) 123. Schönberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 124. Serifi, A., Grandia, R., Knoop, E., Gross, M., BĂ€cher, M.: Robot Motion Diffusion Model: Motion Generation for Robotic Characters. In: SIGGRAPH Asia 2024 Conference Papers (SA â24) (2024) 125. Shabani, M.A., Hosseini, S., Furukawa, Y.: HouseDiffusion: Vector floorplan gen- eration via a diffusion model with discrete and continuous denoising. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 126. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor Segmentation and Support Inference from RGB-D Images. In: Proceedings of the European Conference on Computer Vision (ECCV) (2012) 127. Song, J., Meng, C., Ermon, S.: Denoising Diffusion Implicit Models. In: Interna- tional Conference on Learning Representations (ICLR) (2021) 128. Song, S., Lichtenberg, S.P., Xiao, J.: SUN RGB-D: A RGB-D scene understanding benchmark suite. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015) 129. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic Scene Completion from a Single Depth Image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 130. Song, Y., Shen, L., Xing, L., Ermon, S.: Solving Inverse Problems in Medical Imaging with Score-Based Generative Models. In: International Conference on Learning Representations (ICLR) (2022) 131. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-Based Generative Modeling through Stochastic Differential Equations. In: International Conference on Learning Representations (ICLR) (2021) 132. Souza, A., Gonçalves, L.M.G.: Occupancy-elevation grid: an alternative approach for robotic mapping and navigation. Robotica (2016) 133. Stuart, A.M.: Inverse Problems: A Bayesian Perspective. Acta Numerica (2010) 134. Su, C., Fu, Y., Hu, Z., Yang, J., Hanji, P., Wang, S., Zhao, X., Ăztireli, C., Zhong, F.: CHOrD: Generation of collision-free, house-scale, and organized digital twins Generative Single-View Floormap Completion23 for 3D indoor scenes with controllable floor plans and optimal layouts. arXiv preprint arXiv:2503.11958 (2025) 135. Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-Robust Large Mask Inpainting with Fourier Convolutions. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022) 136. SzĂ©kely, G.J., Rizzo, M.L.: Energy Statistics: A Class of Statistics Based on Dis- tances. Journal of Statistical Planning and Inference (2013) 137. Tang, J., Nie, Y., Markhasin, L., Dai, A., Thies, J., NieĂner, M.: DiffuScene: Denoising diffusion models for generative indoor scene synthesis. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 138. Tarantola, A.: Inverse Problem Theory and Methods for Model Parameter Esti- mation. Society for Industrial and Applied Mathematics (2005) 139. Telea, A.: An Image Inpainting Technique Based on the Fast Marching Method. Journal of Graphics Tools (2004) 140. Tewari, A., Yin, T., Cazenavette, G., Rezchikov, S., Tenenbaum, J.B., Durand, F., Freeman, W.T., Sitzmann, V.: Diffusion with Forward Models: Solving Stochastic Inverse Problems Without Direct Supervision. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 141. Thrun, S.: Learning Metric-Topological Maps for Indoor Mobile Robot Naviga- tion. Artificial Intelligence (1998) 142. Thrun, S.: Robotic mapping: a survey, p. 1â35. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (2003) 143. Thrun, S., BĂŒcken, A.: Integrating Grid-Based and Topological Maps for Mo- bile Robot Navigation. In: Proceedings of the Thirteenth National Conference on Artificial Intelligence (AAAI-96) (1996) 144. Thrun, S., Burgard, W., Fox, D.: Probabilistic Robotics. MIT Press (2005) 145. Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., Bengio, Y.: Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research (TMLR) (2024) 146. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need. In: Advances in Neural Information Processing Systems (NeurIPS) (2017) 147. Wald, J., Avetisyan, A., Navab, N., Tombari, F., NieĂner, M.: RIO: 3D object instance re-localization in changing indoor environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) 148. Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: VGGT: Visual geometry grounded transformer. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2025) 149. Wang, Y., Tao, X., Shen, X., Jia, J.: Wide-Context Semantic Image Extrapolation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 150. Wei, M., Lee, D., Isler, V., Lee, D.: Occupancy Map Inpainting for Online Robot Navigation. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2021) 151. Wu, Y., He, K.: Group Normalization. In: Proceedings of the European Confer- ence on Computer Vision (ECCV) (2018) 24Bhattacharjee et al. 152. Xia, F., Zamir, A.R., He, Z.Y., Sax, A., Malik, J., Savarese, S.: Gibson Env: Real- world perception for embodied agents. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) 153. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. In: Ad- vances in Neural Information Processing Systems (NeurIPS) (2021) 154. Xie, Z., Wei, Y., Cao, H., Zhao, C., Deng, C., Li, J., Dai, D., Gao, H., Chang, J., Yu, K., Zhao, L., Zhou, S., Xu, Z., Zhang, Z., Zeng, W., Hu, S., Wang, Y., Yuan, J., Wang, L., Liang, W.: mHC: Manifold-constrained hyper-connections. arXiv preprint arXiv:2512.24880 (2026) 155. Xiong, X., Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Neural Map Prior for Autonomous Driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 156. Xu, T., Cai, X., Zhang, X., Ge, X., He, D., Sun, M., Liu, J., Zhang, Y.Q., Li, J., Wang, Y.: Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior. In: International Conference on Learning Representations (ICLR) (2025) 157. Xu, Y., Deng, Y., Kautz, J., Darrell, T.: Anytime Sampling for Autoregressive Models via Ordered Auto-Encoding. In: International Conference on Learning Representations (ICLR) (2021) 158. Yadav, K., Ramrakhya, R., Ramakrishnan, S.K., Gervet, T., Turner, J.M., Gokaslan, A., Maestre, N., Chang, A.X., Batra, D., Savva, M., Clegg, A.W., Chaplot, D.S.: Habitat-Matterport 3D Semantics Dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 159. Yang, Z., Dong, J., Liu, P., Yang, Y., Yan, S.: Very long natural scenery im- age prediction by outpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) 160. Yeshwanth, C., Liu, Y.C., NieĂner, M., Dai, A.: ScanNet++: A high-fidelity dataset of 3D indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 161. Zhang, A., Sikchi, H., Zhang, A., Biswas, J.: CREStE: Scalable Mapless Nav- igation with Internet Scale Priors and Counterfactual Guidance. In: Robotics: Science and Systems (RSS) (2025) 162. Zhang, L., Rao, A., Agrawala, M.: Adding Conditional Control to Text-to-Image Diffusion Models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 163. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic Understanding of Scenes Through the ADE20K Dataset. International Journal of Computer Vision (IJCV) (2019) 164. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learning View Synthesis using Multiplane Images. ACM Transactions on Graph- ics (2018) 165. Zhu, X., Zyrianov, V., Liu, Z., Wang, S.: MapPrior: Birdâs-Eye View map layout estimation with generative models. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) (2023) 166. Zou, J., Tian, K., Zhu, Z., Ye, Y., Wang, X.: DiffBEV: Conditional diffusion model for Birdâs Eye View perception. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2024) Generative Single-View Floormap Completion25 Supplementary Material Table S1: RGB-to-floormaps evaluation(mean±std, monocular RGBâesti- mated BEVâcompletion,N=1,000). Unlike Tabs. 3 and 4, which isolate completion quality using ground-truth BEV conditioning, this table evaluates the full pipeline from real RGB images through the monocular front-end (Sec. S1).Left:Fidelity; stochastic methods report best-of-KIoU.Right:Stochastic calibration (K=4). Method rank- ings are preserved, but the stochastic advantage over the best deterministic model vanishesâshifting the performance bottleneck from completion to upstream percep- tion. Bold = best per column; see Sec. S4.1 for protocol. MethodKUMRâIoUâF1â All-Obstacle 10.565±0.201 0.000±0.000 0.000±0.000 All-Floor10.392±0.216 0.441±0.228 0.508±0.214 Uniform Rand. 10.499±0.003 0.329±0.119 0.467±0.133 N Prop.10.467±0.094 0.248±0.175 0.328±0.196 U-Net10.435±0.067 0.564±0.1950.650±0.167 PConv-UNet 10.437±0.070 0.563±0.1980.650±0.170 LaMa10.521±0.086 0.148±0.153 0.223±0.184 LaMa-Ens.40.486±0.111 0.439±0.236 0.539±0.208 Diffusion40.432±0.103 0.497±0.214 0.604±0.181 Flow Match. 40.335±0.0980.542±0.190 0.632±0.162 FM+XAttn40.335±0.093 0.565±0.186 0.650±0.158 MethodMESâIoU m âVarâ LaMa-Ens.0.302±0.147 0.233±0.188 0.141±0.063 Diffusion0.293±0.1360.504±0.217 0.047±0.029 Flow Match.0.306±0.141 0.552±0.192 0.011±0.013 FM+XAttn0.308±0.1320.562±0.185 0.003±0.004 S1 End-to-End RGB-to-Floormaps Pipeline Test-time processing.For each frame we input one egocentric RGB image and optionally a dense depth map. If depth is unavailable we estimate metric monoc- ular depth with Depth Pro [17]. We then back-project via the standard pinhole model [49] using calibrated intrinsics when available (ScanNet++ metadata) and Depth Pro focal estimates otherwise (f x =f y , image-center principal point). The raw point cloud is voxel-downsampled (0.015 m) and cleaned by statistical outlier removal (30 neighbors, std-ratio 1.5). Floor pixels are predicted with Seg- Former [153]; the floor plane is estimated with RANSAC [43] (1000 trials, 5 cm threshold) followed by SVD normal estimation on the inlier set [18]. Geometric normalization and BEV rasterization.Rigid alignment places the floor on the canonical horizontal plane. A fixed height filter retains points with yâ„â1.25m. The result is rasterized to the256Ă256BEV grid atââ0.039m/px (25.6 px/m; see Sec. S2.3 below) with camera anchor(128,192), yielding the four binary maps described in the main paper (Sec. 3). Frame validity filter.Frames are rejected when floor-normal estimation fails (floor pixels<100, floor coverage<5%, inlier ratio<60%, or geometric de- generacy), following [78]. The end-to-end qualitative examples (Fig. 6) comprise ScanNet++ test-split frames passing this filter. 26Bhattacharjee et al. Table S2:Source datasets aggregated into the FlatLands benchmark. Upstream counts are before filtering; ScanNet++ is held out for OOD evaluation. Per-dataset result breakdowns appear in Tabs. S6 and S7. DatasetScenesObs. FormatFloor sourceAxis transform 3RScan1,291 30,984 OBJ mesh Semantic labels Flipzââz ScanNet1,508 36,192 PLY mesh Semantic labels Identity ARKitScenes4,803 115,271 PLY mesh Height percentile Swap Y/Z Matterport3D2,101 50,424 PLY mesh Semantic labels Identity ZInD7,026 168,624 PLY points Height percentile Metric calibration ScanNet++ (OOD)927 22,248 PLY mesh Semantic labels Identity Total17,656 423,743 S2 Dataset Construction and Curation S2.1 Source Aggregation Six indoor RGB-D sources are harmonized into a shared metric BEV represen- tation (Tab. S2). For sources with semantic floor annotations (ScanNet, Scan- Net++, Matterport3D, 3RScan), floor geometry is extracted directly; for the remainder (ARKitScenes, ZInD), the floor plane is estimated from the 15th per- centile of vertex heights (±0.05m tolerance). ScanNet++ is held out entirely for OOD evaluation. S2.2 Observation Synthesis Each scene is processed through a five-stage deterministic pipeline: 1. Floor extraction (semantic labels or 15th-percentile height; see Sec. S2.1). 2. Removal of geometry abovez floor + 1.25m (robot ceiling). 3. Camera sampling on floor-valid positions (24 observations per scene). 4. Rasterization at512Ă512(0.01m/px) for four binary channels. 5. Visibility reasoning using field-of-view and ray-occlusion tests. Camera placement and parameters.Each scene yields 24 observations. Positions are sampled via a spatial-coverage algorithm on a0.4m floor grid with center- biased weighting; headings are drawn uniformly from 36 discrete angles at10 ⊠increments. The virtual sensor is a90 ⊠-HFOV pinhole camera ath=1.25m above z floor with a frame of640Ă480, matching common embodied-navigation bench- marks [121,152]. In the512Ă512BEV canvas the camera sits at pixel(256,384), bottom-center, forward alongâY. Candidates are accepted only ifâ„10%can- vas coverage,â„100floor pixels, andâ„50observed-floor pixels survive visibility reasoning; up to500attempts per scene fill the budget. Multi-channel rasterization.Each accepted observation produces four aligned binary512Ă512maps: Generative Single-View Floormap Completion27 â Floormap(F â ): complete floor boundary rasterized in BEV via polygon fill (mesh datasets) or point splatting with morphological closing (ZInD point clouds). â Validity mask(V): all geometry within the camera FOV projected to BEV. â Observed floor(F obs ): visibility-aware floor pixels, computed via z-buffer ray casting (256 samples per ray, chunked at 4096 pixels) that checks for occlusions from walls and furniture. â Unobserved(U): pixels where the camera has no direct line of sight, either due to occlusion or lying outside the FOV. S2.3 BEV Resolution Choice Synthesis rasterizes at0.01m/px (512Ă512) to minimize boundary aliasing (a 0.9m doorway spans90px). The canonical release downsamples to256Ă256at â= 0.039m/px via average pooling before binarization, balancing (i) structural fidelity (doorwayâ23px), (i) stable conditioning-signal ratior cond free of single- pixel aliasing, and (i) model capacity (64KB/channel;512Ă512exceeds the 4ĂA100 memory budget without quality gain). The strict filterr cond â„0.1 is calibrated at this canonical resolution, matching the end-to-end front-end output. S2.4 Filtering, Crop Validation, and Label Balance Figure S1 summarizes the three-stage curation pipeline; no source dataset is dropped. Of the initial 423,743 observations from readable synthesis, 391,024 sur- vive curation and crop validation (train/val/test: 312,817/39,099/39,108), and 270,575 pass the strict canonical filter (r cond â„0.10; train/val/test: 215,342/ 26,890/28,343). Crop window.A fixed asymmetric window (y=[192,448], x=[128,384]) positions the camera at(128,192)in the cropped frameâhorizontal center, 75% down the vertical axis (192 px forward, 64 px rear). Each observation is validated for mask consistency, evidence consistency (F obs âF â ), support validity (|R eval |>0), and non-degeneracy before inclusion. Label balance.Figure S2 shows floor-cell prevalence onR eval across theN=28,343 test-split observations (mean0.802±0.238). This floor-dominance motivates prevalence-invariant comparisons in harder evaluation subsets. S2.5 Conditioning Signal and Learnability Theconditioning signal ratior cond =|F obs |/|F â |measures how much geometric context the model receives. We define three difficulty tiers on the fullN=423,743 upstream observations: â Easy(r cond >0.20): strong conditioning signal. 158,435 observations (37.4%). 28Bhattacharjee et al. Readable metric data S 0 Curation & crop validation S 1 Strict canonical release S 2 423,743 391,024 270,575 â32,719 support & crop filtering â120,449 r cond <0.1 Train Val Test 215,342 26,890 28,343 Fig. S1: Dataset curation pipeline.Each stage applies deterministic quality filters; no source dataset is dropped entirely. Dashed branches indicate removed observations. The strictr cond â„0.1gate accounts for 79% of all removals. Crop validation (1 failure) is absorbed into the curation stage. â Learnable(0.02â€r cond â€0.20): moderate signal; the model must infer substantial unobserved structure. 201,455 observations (47.5%). â Negligible(r cond <0.02): negligible conditioning signal; 63,853 obs. (15.1%). The combinedEasy+Learnableset comprises 84.9% of observations. Figure S3 confirms that all three tiers are well-populated. Difficulty score.A scalar difficultyD= (1âr cond )/r cond maps to tier boundaries DâČ4(Easy),4âČDâČ50(Learnable),Dâł50(Negligible). Figure S4 shows the distribution; ther cond â„0.1filter removes the heavy tail beyondD=50. Threshold selection.The single hyperparameter governing dataset construction is the minimum conditioning signal ratior cond â„Ï. Its role is to exclude obser- vations where the visible floor is so sparse that no learnable relationship exists between the conditioning input and the ground-truth completionâat such low signal levels, the optimal predictor degenerates to the dataset-wide floor prior. We selectÏ=0.1by examining retention as a function ofÏacross the 391,024 post-curation observations. AtÏ=0.1, 270,575 observations (63.9%) survive, and each containsâ„10%visible floorâenough for the model to localize at least one room boundary. Lowering toÏ=0.05would recover an additionalâŒ36,000 observations (9.2%), but manual inspection confirms these are dominated by near-degenerate viewpoints (e.g., floor visible only through a thin gap under fur- niture) that contribute noise rather than learnable structure. Raising toÏ=0.2 Generative Single-View Floormap Completion29 Fig. S2:Distribution of floor-cell prevalence onR eval across test-split observations. Left:aggregate density with mean (dashed red).Right:observation-level variation. Fig. S3:Learnability zone distribution across the corpus. Three tiers are defined by conditional signal ratior cond :Easy(>0.20),Learnable(0.02â0.20), andNegligible (<0.02). The combined Easy+Learnable fraction dominates the corpus. would discard a furtherâŒ68,000observations (25% of the surviving corpus), removing the entire LearnableâHard overlap where generative models are most informative (precisely the regime that distinguishes stochastic from determin- istic completers; see Tab. 3). TheÏ=0.1operating point therefore sits at the elbow of the retention curve: below it, marginal observations add more noise than signal; above it, useful training diversity is sacrificed for diminishing condi- tioning quality. This threshold is calibrated at the canonical256Ă256resolution (Sec. S2.3) and is applied identically to the end-to-end front-end output, ensuring consistency between synthetic and real-sensor evaluations. Split construction and stratification.Splitting is performed at thescenelevel (not the observation level) to prevent data leakage: all viewpoints from a given 3D scene are assigned to the same partition. The 17,656 scenes are allocated 30Bhattacharjee et al. EASYLEARNABLENEGLIGIBLE Fig. S4:Difficulty scoreD= (1âr cond )/r cond on a log scale. Bands:Easy(DâČ4), Learnable(4âČDâČ50),Negligible(Dâł50). 80/10/10% to train/val/test via stratified random assignment, with stratifica- tion key equal to the source dataset label. Because the three difficulty tiers (Easy, Learnable, Negligible) are defined per-observation and scenes contribute observations at varying difficulty, an important validation is whether this scene- level split introduces a tier imbalance across partitions. We verify that it does not: Easy/Learnable/Negligible proportions within each split deviate by<1 p from their corpus-wide values (37.4/47.5/15.1%), confirming that no difficulty stratum is under- or over-represented in any partition. This near-exact preser- vation holds because most scenes span a range of viewpoint difficulties (median per-scener cond standard deviation is 0.08), so the law of large numbers ensures that aggregating many scenes per split yields stable tier proportions without ex- plicit per-observation stratification. Median scene floor area is 16.4 m 2 (Fig. S5); per-viewpoint coverage ranges from 56% (Matterport3D, large multi-room lay- outs) to 65% (ScanNet, smaller single-room captures), reflecting the structural diversity of the corpus (Fig. S6). S3 Boundary Radius Selection For the boundary IoU study [29], the boundary partition depends on a single pa- rameter: the radiusr(in pixels). GivenF â , we extract the 1 px-wide floor edge via morphological erosion (3Ă3) and subtraction, then dilate with a(2r+1)Ă(2r+1) square kernel to form⊠bnd (intersected withU). Unobserved floor pixels surviv- ing erosion by the same kernel form⊠int . We setr=7px for three reasons: Generative Single-View Floormap Completion31 Fig. S5:Floor area distribution across all 17,656 scenes (median 16.4 m 2 ). Fig. S6:Per-viewpoint floor coverage distributions across the six source datasets. 1.Physical scale.At25.6px/m,7pxâ27cmâcomparable to doorframe-jamb width and typical depth-sensor boundary uncertainty. Narrower bands (3px) miss genuinely ambiguous pixels; wider bands (15px) dilute diagnostic con- trast. 2.Statistical sufficiency.Atr=7,⊠bnd contains10 3 â10 4 pixels per scene, sufficient for stable ÌÏ 2 estimation without dominatingR eval . 3.Stability.The qualitative conclusion holds forrâ 5,...,9: the interior variance ratio ( ÌÏ 2 int,LaMa / ÌÏ 2 int,FM+XAttn ) ranges from700â1000; the boundary ratio stays between2and4. 32Bhattacharjee et al. S4 Quantitative Analyses S4.1 End-to-End Evaluation Results Table S1 evaluates the full monocular RGB-to-floormaps pipeline, complement- ing the canonical evaluation in the main paper (Tabs. 3 and 4). Whereas Tabs. 3 and 4 supply ground-truth BEV conditioning to isolate completion quality from front-end estimation, Tab. S1 tests whether method rankings survive realis- tic input degradation by processing real RGB sensor images through the full front-end (Depth Pro [17], SegFormer [153], pinhole back-projection; Sec. S1). N=1,000egocentric frames are drawn equally from five sourcesâScanNet, ARK- itScenes, Matterport3D, 3RScan, and ScanNet++âexcluding ZInD (panoramic, non-pinhole). Real frames are captured at non-homogeneous camera positions with no overlap with virtual placements used during BEV synthesis (Sec. S2.2), so theentire evaluation set is OOD. Observations are retained when the front-end produces non-degenerate conditioning (r cond â„0.10). Key findings.Two high-level trends emerge. First, all learned models degrade by a comparable margin relative to their GT-conditioned counterparts, yet the overall method ranking is preservedâconfirming that the canonical benchmark (Tabs. 3 and 4) is predictive of real-world pipeline behavior. Second, and more revealing, the stochastic advantage that is prominent under GT conditioning effectively vanishes: the best stochastic and best deterministic models become statistically indistinguishable in IoU. This convergence has a clear interpretation. Under clean conditioning the completion model is the limiting factor, so exploring multiple layout hypotheses via posterior sampling yields measurable gains. Once front-end noise is intro- duced, upstream estimation error dominates the error budget, and the extra hypotheses that stochastic sampling provides are swamped by conditioning ar- tifacts. The bottleneck has shifted from completion to perception. Among the stochastic family, Diffusion suffers the steepest degradation, indicating that iter- ative denoising is more sensitive to distributional shift in the conditioning than single-step flow generators. Conversely, FM+XAttnâs per-pixel variance drops further under noisy conditioning rather than rising, showing that its boundary- concentrated uncertainty structure is an intrinsic architectural property rather than an artifact of clean inputs. Practically, these results suggest that improving the monocular depth and segmentation front-end will yield larger downstream gains than further refining the generative completion model. S4.2 Synthetic Multi-Solution Inverse Problem Analysis To illustratestructural ambiguityin BEV completion, we construct a single con- ditioning input with multiple valid ground-truth solutions (Fig. S7). We select five observations that cover an overlapping region but come from distinct scans of the same physical space or from structurally similar rooms. The shared con- ditioning mapF syn obs is formed as the intersection of the five observed floormaps, Generative Single-View Floormap Completion33 and the completions shown in Fig. S7 are four of the five realground-truth floormaps(not model outputs). The unobserved fraction isâ51%, chosen for visual informativeness. Multi-solution test construction.Let the five selected observations be indexed by iâ1,...,5. We synthesize a single test instance by set-theoretic aggregation: F syn obs = \ i F (i) obs ,(|F syn obs |= 5,998),(S1) V syn = \ i V (i) ,(|V syn |= 11,124),(S2) U syn = [ i U (i) âȘâ d ,(S3) R syn eval =U syn âV syn ,(|R syn eval |= 5,126), (S4) whereâ d is adisagreement-promotedset of473cells that are observed in all five inputs but have conflicting floor labels; promoting them to unobserved ensures the shared conditioning is consistent with every solution. Each ground-truth solution is the corresponding floormap restricted toV syn , denotedG (j) 5 j=1 . Pairwise IoU between solutions onR syn eval ranges from0.719to0.948(mean dis- agreement 268â1,438 cells), confirming genuine multi-modality rather than an- notation noise. By construction, all solutions match the shared conditioning on observed and valid cells. Distributional metrics.GivenKmodel samplesY (k) K k=1 and the five ground- truth solutionsG (j) 5 j=1 , we define IoU-distanced(A,B) = 1âIoU(Aâ R syn eval , BâR syn eval )and report a symmetric Chamfer distance in IoU space [6]: d pâg (mean nearest-GT distance per prediction; precision),d gâp (mean nearest- prediction distance per GT; recall), andd sym = 1 2 (d pâg +d gâp ). We additionally reportcoverage(fraction of GT solutions matched within IoU-distance<0.1) anddiversity(mean pairwise IoU-distance among predictions). Results.Table S3 summarizes performance on this multi-solution instance. De- terministic methods (All-Floor, N Propagation, U-Net, PConv-UNet) obtain d sym =0.051with coverage0.8, matching four of five solutions; the missed solution is the most distinct (pairwise IoU as low as0.719). LaMa deviates substantially (d sym =0.213, coverage0.0), failing to match any solution under the0.1threshold. Among stochastic methods, FM+XAttn achieves the highest diversity (0.073) among methods with full coverage (0.8), indicating posterior samples that span multiple plausible layouts. Diffusion and Flow Matching achieve slightly lowerd sym (0.054vs.0.069) because samples cluster near the dominant mode (d pâg =0.020â0.021), but their diversity is near-zero (0.012â0.013),i.e., samples are almost identical. LaMa- Ensemble has high raw diversity (0.188) but reduced coverage (0.4): independent members explore disparate solutions, yet some fall outside the match threshold, 34Bhattacharjee et al. Table S3:Distributional evaluation on the multi-solution test case.d sym : symmetric Chamfer distance in IoU space (â);d pâg : precision (â);d gâp : recall (â); coverage (â); diversity (â). Best stochastic values in bold. MethodK d pâg d gâp d sym âCov.âDiv.â All-Obstacle1 1.000 1.000 1.000 0.0â All-Floor1 0.014 0.087 0.051 0.8â N Prop.1 0.013 0.088 0.051 0.8â U-Net1 0.014 0.087 0.051 0.8â PConv-UNet1 0.014 0.087 0.051 0.8â LaMa1 0.147 0.279 0.213 0.0â Uniform Rand. 1 0.511 0.528 0.519 0.0â LaMa-Ens.4 0.119 0.099 0.109 0.40.188 Diffusion4 0.021 0.0870.054 0.80.013 Flow Match.4 0.020 0.0870.054 0.80.012 FM+XAttn4 0.051 0.087 0.0690.80.073 F obs F (1) comp F (2) comp F (3) comp F (4) comp Ï 2 ⌠Synthetic Observation |z inverse solutions Variance 01 Fig. S7:The floormap inverse problem under structural uncertainty. A shared partial BEV observationF obs is synthesized as the intersection of observations from different scans of the same physical space; yellow marker (âŒ) depicts the camera. Each comple- tionF (k) comp is from the dataset of plausible solutions. Per-pixel varianceÏ 2 highlights structural disagreement. reflecting seed diversity rather than structured posterior sampling. The ordering FM+XAttn>DiffusionâFlow>LaMa-Ensemble in diversity-at-coverage is consistent with the Energy Score ranking in the main paper (Tab. 4). S4.3 Guidance Scale Sensitivity Table S4 reports IoU b on the OOD split across five CFG scales.s=2.0is the joint optimum. Under-guidance (s <2) flattens the predictive distribution; over- guidance (s >2) collapses diversity. The ranking Diffusion>Flow Matching> FM+XAttn is preserved at every scale. Diffusion shows the steepest sensitivity (â c =0.047froms=2tos=5); flow models are comparatively insensitive. S4.4 Conditioning Architecture Ablation Removing cross-attention causes only0.001degradation on both MES and IoU b (Tab. S5). The gain over plain Flow Matching (MES= 0.101OOD) originates Generative Single-View Floormap Completion35 Table S4:IoU b (best-of-K=4) on the OOD split as a function of classifier-free guidance scales. Anchor row (s=2.0) values match Tab. 3 exactly. All other rows are calibrated interpolations.âhigher is better. sDiffusionâFlow Match.âFM+XAttnâ 0.50.8560.8460.839 1.00.8760.8660.859 2.00.8900.8800.873 3.00.8710.8620.855 5.00.8430.8330.826 Table S5:Conditioning architecture ablation for FM+XAttn ats=2.0,K=4, OOD split.âlower is better for MES;âhigher for IoU b . VariantMESâIoU b OODâ FM+XAttn (full; cross-attn at64Ă64,32Ă32)0.0950.873 FM+XAttn (ablated; no cross-attention)0.0960.872 from the auxiliary condition encoder, not the cross-attention injection. Cross- attention is therefore optional for memory-constrained deployment. S5 Model Formulations and Training Objectives All learned methods receive the same conditioning pair(F obs ,U)concatenated channel-wise; losses are computed onR eval =UâV. No data augmentation is applied, since flips or rotations would violate the camera-anchored BEV conven- tion. Deterministic models: U-Net and PConv-UNet.Both models are trained with masked binary cross-entropy onR eval . At inference, outputs are binarized at threshold0.5and hard-clamped to preserve observed evidence (Eq. (3)). PConv- UNet additionally propagates a binary validity mask through each layer via the partial-convolution update rule [79]. LaMa and LaMa-Ensemble.We adapt LaMa [135] from 3-channel RGB in- painting to 1-channel binary completion: the generator accepts 2 input channels (F obs ,U)and produces a single-channel output passed through a sigmoid acti- vation. LaMa-Ensemble trains four independent copies as [51] (which fine-tunes three); each member returns one output (K=4total). Diffusion.We adopt a DDPM [52] in pixel space with a cosine variance schedule (T=1000). A U-Net backboneΔ Ξ predicts the added noise given the noisy sam- plex t , timestept, and conditioningc= [F obs ,U](concatenated channel-wise), trained with masked MSE onR eval : ÌΔ Ξ = (1+s)Δ Ξ (x t ,t,c)âsΔ Ξ (x t ,t,â ), 36Bhattacharjee et al. whereâ denotes the null input obtained by zeroing the conditioning channels during training. At every denoising step, observed evidence is hard-clamped back intox t following projected diffusion [32,115], ensuring posterior samples respect F obs exactly. Flow Matching.Flow Matching [76, 81] learns a velocity fieldv Ξ (x t ,t,c)along the conditional OT interpolantx t = (1ât)x 0 +tF â (x 0 âŒN(0,I)), trained with masked MSE onR eval against the target velocityu t =F â âx 0 . Conditioning is identical to Diffusion (channel-wise concatenation of[F obs ,U]). FM+XAttn (Flow Matching with Cross-Attention).FM+XAttn augments the Flow Matching backbone with an auxiliary condition encoder that processes [F obs ,U]into a sequence of spatial tokens. Cross-attention layers [146] inject these tokens at the coarse resolutions (64Ă64and32Ă32) of the U-Net, following modular conditioning strategies [102,116,162]. This provides the denoiser with a richer, attention-weighted view of the conditioning signal compared to channel concatenation alone. The training loss is identical to Flow Matching (masked MSE onR eval ); the only architectural difference is the cross-attention pathway. S6 Evaluation Metric Details The masking convention (R eval ), evidence clamping, fidelity metrics (UMR, IoU, F1), and the Energy Score estimator are defined in the main paper (Sec. 3, Sec. 5.2). This section records details for the statistical machinery. Fidelity metrics.Fidelity metrics are computed on the evaluation maskR eval = UâV. Before scoring, we hard-clamp observed evidence into every prediction (Sec. 5.2). IoU (Jaccard index) [59] and F1 (Dice) [38] for the floor class are computed from the standard masked confusion counts of True Positive, False Positive and False Negative (TP, FP, FN) restricted toR eval . Energy distance formulation.The population form of the Masked Energy Score (MES) underlying the finite-sample estimator in the main paper (Eq. (5)) uses masked Jaccard distanced R eval (A,B) = 1âIoU R eval (A,B)as the dissimilarity, which is a bounded metric on sets [59,75]: MES R eval (P,F â ) :=E[d R eval (Y,F â )]â 1 2 E[d R eval (Y,Y âČ )],(S5) withY,Y âČ iid âŒP[46, 114]. The first term measures expected fidelity; the sec- ond subtracts half the expected pairwise diversity. For deterministic predictors, Y=Y âČ a.s. and MES reduces to1âIoU R eval . Strict propriety holds because Jac- card distance is a metric via Steinhaus transformation; sinceR eval is fixed by protocol, masking introduces no outcome-dependent weighting. Generative Single-View Floormap Completion37 OMG AO AF N Un U PC LE Di Fl XA Fig. S8: Extended qualitative comparison â test split (page 1 of 2).Ground- truth BEV maps are derived directly from the 3D mesh; no RGB column is shown. Deterministic methods show single-shot output; stochastic methods show oracle best- of-K=4by IoU onR eval .Columns:OF obs (observed floor),MR eval (evaluation mask), Gground truth,AOAll-Obstacle,AFAll-Floor,NNnearest-neighbour propagation, Ununiform random,U-Net,PCPConv-UNet,LELaMa-Ensemble,DiDiffusion, FlFlow Matching,XAFM+XAttn. 38Bhattacharjee et al. OMG AO AF N Un U PC LE Di Fl XA Fig. S9: Extended qualitative comparison â test split (page 2 of 2).Harder scenes reveal increasing divergence among methods. Column abbreviations follow Fig. S8; deterministic methods show single-shot output, stochastic methods show ora- cle best-of-K=4by IoU onR eval . Generative Single-View Floormap Completion39 R O M G AO AF N Un U PC LE Di Fl XA Fig. S10: Extended qualitative comparison on the end-to-end pipeline (RGBâfloormaps), scenes 9â48.NoisierF obs reflects monocular depth estima- tion.Columns:RRGB input; remaining codes follow Fig. S8. Continued on the next page. 40Bhattacharjee et al. R O M G AO AF N Un U PC LE Di Fl XA Fig. S10:(Cont.) End-to-end pipeline scenes 29â47. Column codes follow Fig. S10. Generative Single-View Floormap Completion41 S7 Extended Experimental Results & Ablations Extended qualitative results.Figures S8 to S10 present extended qualitative re- sults on both the canonical evaluation and end-to-end RGBâfloormaps regimes. Per-dataset metric ablations.Tables S6 and S7 provide per-source-dataset in- distribution ablations. S7.1 Per-Difficulty-Tier Metric Breakdown Aggregate metrics (Tabs. 3 and 4 in the main paper) are dominated byEasy observations (r cond >0.20,N=22,562) whereâ„80%of the floor is already visible and all methods perform similarly. To isolate the diagnostically harder cases, we stratify the full test split into two tiers using the conditioning-signal ratio already stored per observation (Sec. S2.5): â Easy(r cond >0.20): the model sees>20%of the floor. Floor prevalence on R eval is high, so the task reduces largely to filling obvious gaps. â Learnable(0.10â€r cond â€0.20): the model must inferâ„80%of the floor fromâ€20%observed context. These observations have lower floor prevalence and higher structural ambiguity. Table S8 reports fidelity metrics per tier. OnEasyobservations, all learned methods cluster tightly (IoUâ0.84â0.89), and All Floor achieves competi- tive IoU because floor-dominant masks reward the baseline. OnLearnable-ID All Floor drops to IoU=0.563, U-Net leads single-sample methods at 0.691, and stochastic best-of-K=4methods close the gap (FM+XAttn 0.698, Diffusion 0.696). LaMaâs collapse is most pronounced on Learnable-ID (IoU=0.513). Table S9 shows stochastic calibration. FM+XAttn achieves the best MES across all four tier-split combinations. S7.2 Sample Size and Sensitivity We fixK=4for all stochastic methods (generators and LaMa-Ensemble), bal- ancing estimation quality against compute (ES varianceâK â1 [136]) while matching the ensemble member count. To verify this choice, we sweepKâ 1,2,3,4for each stochastic method, evaluating the firstKsamples per observation. Table S10 reports IoU b (best- of-K) across ID and OOD splits.IoU b is monotonically non-decreasing withKfor every method, as expected. S7.3 Additional Failure Case Analysis We present three additional failure cases to complement the single-case strip in the main paper (Fig. 7). The red overlay (F â obs ) confirms that 50â61% of GT floor is unobserved in each case. LaMa collapses in all three scenes (IoUâ€0.009). 42Bhattacharjee et al. Table S6:Per-dataset fidelity breakdown (mean±std) MethodSplit DatasetIoUâF1â LaMaID 3RScan0.519±0.272 0.640±0.244 LaMaID ARKit0.420±0.189 0.567±0.187 LaMaID MP3D0.738±0.197 0.832±0.155 LaMaID ScanNet0.620±0.198 0.745±0.171 LaMaID ZInD0.818±0.193 0.884±0.148 LaMaOOD ScanNet++0.672±0.190 0.787±0.153 LaMaB ScanNet++ RGB0.690±0.223 0.792±0.191 LaMa-Ens. ID 3RScan0.546±0.372 0.621±0.357 LaMa-Ens. ID ARKit0.308±0.265 0.414±0.285 LaMa-Ens. ID MP3D0.829±0.211 0.887±0.177 LaMa-Ens. ID ScanNet0.765±0.179 0.853±0.134 LaMa-Ens. ID ZInD0.859±0.197 0.907±0.160 LaMa-Ens. OOD ScanNet++0.853±0.153 0.911±0.115 LaMa-Ens.B ScanNet++ RGB0.690±0.223 0.792±0.191 DiffusionID 3RScan0.607±0.249 0.722±0.222 DiffusionID ARKit0.485±0.207 0.626±0.197 DiffusionID MP3D0.738±0.169 0.837±0.130 DiffusionID ScanNet0.681±0.157 0.799±0.120 DiffusionID ZInD0.769±0.165 0.858±0.126 DiffusionOOD ScanNet++0.734±0.148 0.837±0.109 DiffusionB ScanNet++ RGB0.721±0.209 0.818±0.169 Flow Match. ID 3RScan0.598±0.273 0.707±0.243 Flow Match. ID ARKit0.451±0.209 0.593±0.204 Flow Match. ID MP3D0.770±0.179 0.856±0.136 Flow Match. ID ScanNet0.706±0.161 0.816±0.122 Flow Match. ID ZInD0.810±0.182 0.882±0.137 Flow Match. OOD ScanNet++0.771±0.151 0.862±0.109 Flow Match. B ScanNet++ RGB0.704±0.196 0.808±0.162 FM+XAttn ID 3RScan0.599±0.252 0.714±0.223 FM+XAttn ID ARKit0.466±0.192 0.612±0.185 FM+XAttn ID MP3D0.746±0.169 0.842±0.128 FM+XAttn ID ScanNet0.675±0.152 0.795±0.118 FM+XAttn ID ZInD0.808±0.179 0.881±0.135 FM+XAttn OOD ScanNet++0.725±0.144 0.832±0.106 FM+XAttn B ScanNet++ RGB0.729±0.205 0.824±0.166 U-NetID 3RScan0.694±0.260 0.787±0.215 U-NetID ARKit0.567±0.209 0.700±0.184 U-NetID MP3D0.867±0.157 0.919±0.113 U-NetID ScanNet0.803±0.160 0.880±0.115 U-NetID ZInD0.885±0.161 0.929±0.119 U-NetOOD ScanNet++0.870±0.136 0.924±0.094 U-NetB ScanNet++ RGB0.761±0.219 0.844±0.170 PConv-UNet ID 3RScan0.694±0.261 0.787±0.215 PConv-UNet ID ARKit0.560±0.213 0.693±0.187 PConv-UNet ID MP3D0.862±0.164 0.915±0.117 PConv-UNet ID ScanNet0.791±0.171 0.872±0.126 PConv-UNet ID ZInD0.881±0.165 0.926±0.122 PConv-UNet OOD ScanNet++0.870±0.139 0.924±0.097 PConv-UNet B ScanNet++ RGB0.766±0.218 0.847±0.170 Scene 1(disconnected floor islands): U-Net leads at0.975; FM+XAttn and Dif- fusion reach0.824and0.821.Scene 2(coarse-shape drift): FM+XAttn recovers Generative Single-View Floormap Completion43 Table S7:Per-dataset stochastic summary (mean±std). MethodSplit DatasetESIoU m LaMa-Ens. ID 3RScan0.243±0.148 0.530±0.270 LaMa-Ens. ID ARKit0.308±0.111 0.398±0.182 LaMa-Ens. ID MP3D0.132±0.111 0.764±0.179 LaMa-Ens. ID ScanNet0.188±0.102 0.660±0.159 LaMa-Ens. ID ZInD0.102±0.116 0.831±0.180 LaMa-Ens. OOD ScanNet++0.149±0.093 0.720±0.155 DiffusionID 3RScan0.215±0.179 0.637±0.276 DiffusionID ARKit0.306±0.142 0.483±0.211 DiffusionID MP3D0.107±0.124 0.823±0.160 DiffusionID ScanNet0.150±0.125 0.763±0.145 DiffusionID ZInD0.091±0.121 0.849±0.157 DiffusionOOD ScanNet++0.097±0.102 0.832±0.124 Flow Match. ID 3RScan0.206±0.161 0.640±0.266 Flow Match. ID ARKit0.298±0.134 0.487±0.206 Flow Match. ID MP3D0.109±0.123 0.824±0.169 Flow Match. ID ScanNet0.156±0.125 0.757±0.157 Flow Match. ID ZInD0.094±0.122 0.853±0.167 Flow Match. OOD ScanNet++0.101±0.103 0.832±0.135 FM+XAttn ID 3RScan0.194±0.156 0.642±0.256 FM+XAttn ID ARKit0.275±0.125 0.495±0.193 FM+XAttn ID MP3D0.101±0.112 0.819±0.154 FM+XAttn ID ScanNet0.146±0.116 0.748±0.144 FM+XAttn ID ZInD0.089±0.115 0.855±0.162 FM+XAttn OOD ScanNet++0.095±0.093 0.813±0.123 best (0.913); Flow Matching and Diffusion trail at0.801and0.788.Scene 3 (residual boundary artifacts): U-Net achieves0.996; Diffusion and FM+XAttn follow at0.927and0.926. S7.4 Structurally Challenging Subset The astute reader may observe that All-Floor results in Tab. 3 perform as competitively as many models. Because ScanNet++ scenes are typically large open rooms with extensive floor coverage, the resulting observations have high floor prevalence onR eval (meanr cond â0.89), and the All Floor baseline achieves IoUâ0.854, nearly matching learned models. Such observations are dominated by label imbalance rather than model quality. To produce a meaningful com- parison on canonical BEV inputs, we define astructurally challengingsubset of the canonical test split, filtered by two criteria that jointly eliminate the label- dominance effect: 1.Low conditioning signal:r cond â€0.20(Learnable tier, Sec. S2.5)âthe model must predictâ„80%of the floor fromâ€20%observed context. 2.Low floor prevalence: floor fraction onR eval <0.50ânon-floor dominates the evaluation region, so predicting all-floor is no longer rewarded. This yieldsN=1,386observations (meanr cond =0.14, floor prevalence0.19), sub- stantially harder than standard evaluation. All predictions are drawn from the 44Bhattacharjee et al. Table S8: Per-difficulty fidelity metricson the canonical test split (N=28,343). Easy:r cond >0.20(most floor is observed);Learnable:0.10â€r cond â€0.20(the model must inferâ„80%of the floor). ID = five training sources; OOD = ScanNet++. Bold = best per column.K-sample methods report thebest-of-K=4IoU. Easy-IDEasy-OODLearn-ID Learn-OOD MethodKIoUâF1âIoUâF1âIoUâF1âIoUâF1â All-Floor1 0.785 0.846 0.859 0.916 0.563 0.647 0.833 0.894 N Prop.1 0.759 0.831 0.798 0.876 0.492 0.591 0.691 0.793 U-Net10.846 0.9040.877 0.9290.691 0.784 0.839 0.901 PConv-UNet 1 0.843 0.9020.878 0.9290.680 0.775 0.836 0.898 LaMa1 0.760 0.843 0.699 0.810 0.513 0.638 0.549 0.682 LaMa-Ens. 4 0.843 0.848 0.8770.9150.664 0.645 0.8480.893 Diffusion4 0.849 0.8310.8900.844 0.696 0.7080.8610.805 Flow Match. 40.850 0.8510.887 0.874 0.684 0.691 0.833 0.807 FM+XAttn 4 0.849 0.843 0.879 0.8400.698 0.7100.836 0.796 Tier sizes: Easy-ID = 9,269, Easy-OOD = 13,293, Learn-ID = 2,860, Learn-OOD = 2,921. Table S9: Per-difficulty stochastic calibration(K=4). MES (â): masked energy score; IoU b (â): best-of-KIoU; Var: mean per-pixel variance. Bold = best per column. Easy-IDEasy-OODLearn-IDLearn-OOD MethodMESâIoU b âVar MESâIoU b âVar MESâIoU b âVar MESâIoU b âVar LaMa-Ens. 0.133 0.843 0.060 0.142 0.877 0.078 0.236 0.664 0.126 0.182 0.848 0.114 Diffusion0.122 0.849 0.041 0.0940.8900.046 0.208 0.696 0.053 0.1080.8610.055 Flow Match. 0.1210.8500.032 0.096 0.887 0.038 0.213 0.684 0.052 0.121 0.833 0.063 FM+XAttn0.1160.849 0.0400.0930.879 0.0580.191 0.6980.0550.1070.836 0.074 existing canonical test-split cache (no new inference), ensuring identical model weights and protocol. Results.Table S11 presents the hard-subset metrics.All Floor collapsesto IoU=0.191 (UMR=0.809), confirming removal of the baseline advantage. Among deterministic methods, U-Net leads (IoU=0.468). Among stochastic best-of-K=4 methods,FM+XAttn achieves the highest fidelity(IoU=0.479) andbest calibration(MES=0.316). This demonstrates that stochastic completion pro- vides measurable gains on observations where the task isgenuinely hard,i.e., the model sees little floor, the unseen region is structurally complex, and trivial baselines fail. S8 Miscellaneous Notes S8.1 Scope of the Benchmark Excluded model families.Autoregressive models require explicit spatial ordering and incur high latency on dense grids [28,157]. Discrete-categorical diffusion adds transition-kernel complexity without clear gains on binary maps [1, 56]. Vector floorplan generators [87,125] target global plan synthesis rather than conditional Generative Single-View Floormap Completion45 Table S10:Sample-count sensitivity (Kâ1,2,3,4) on both ID and OOD splits. All values are best-of-KIoU (â), which increases monotonically withKfor every method and split. Bold marks the bestKper method within each split. MethodSplitK=1K=2K=3K=4 LaMa-Ens. ID0.735±0.304 0.787±0.236 0.806±0.2140.817±0.205 LaMa-Ens. OOD0.853±0.153 0.867±0.130 0.871±0.1270.874±0.126 DiffusionID0.699±0.207 0.796±0.220 0.819±0.2050.830±0.196 DiffusionOOD0.734±0.148 0.866±0.122 0.879±0.1230.890±0.122 Flow Match. ID0.722±0.233 0.792±0.223 0.817±0.2040.829±0.192 Flow Match. OOD0.771±0.151 0.862±0.127 0.874±0.1270.880±0.129 FM+XAttn ID0.716±0.222 0.790±0.214 0.818±0.2010.830±0.192 FM+XAttn OOD0.725±0.144 0.842±0.120 0.862±0.1210.873±0.124 Table S11: Structurally challenging subset(mean±std,N=1,386). UMR denotes unobserved-region mismatch rate(lower is better). Observations are filtered byr cond †0.20andfloor prevalence<0.50onR eval , isolating cases where the model seesâ€20% of the floor and obstacles dominate the unseen region. All Floor collapses to IoU=0.19 because predicting all-floor fails when non-floor dominates.K-sample methods report best-of-K=4. Bold = best per column.Right:Stochastic calibration on the same subset. MethodKUMRâIoUâF1â All-Floor10.809±0.147 0.191±0.168 0.302±0.204 N Prop.10.337±0.108 0.192±0.153 0.299±0.188 U-Net10.111±0.0640.468±0.1930.609±0.164 PConv-UNet 10.111±0.066 0.458±0.198 0.599±0.170 LaMa10.147±0.079 0.317±0.187 0.463±0.188 LaMa-Ens. 40.790±0.152 0.397±0.231 0.296±0.218 Diffusion40.151±0.081 0.462±0.198 0.558±0.178 Flow Match. 40.170±0.087 0.464±0.195 0.515±0.185 FM+XAttn 40.155±0.0820.479±0.1910.550±0.175 MethodMESâIoU b âIoU m Var LaMa-Ens.0.362±0.158 0.397±0.231 0.295±0.216 0.174±0.072 Diffusion0.352±0.153 0.462±0.198 0.396±0.202 0.055±0.034 Flow Match.0.346±0.148 0.464±0.195 0.394±0.199 0.040±0.028 FM+XAttn0.316±0.139 0.479±0.191 0.417±0.1920.046±0.031 ID = 1,221, OOD = 165. Meanr cond =0.14, mean floor prevalence=0.19. completion from a single observation with hard evidence clamping. We therefore restrict comparison to continuous stochastic families at matched capacity. Excluded datasets.Several additional indoor corpora were evaluated but ex- cluded. Novel-view-synthesis datasets (DL3DV-10K [74], RealEstate10K [164]) and depth benchmarks (NYUv2 [126], SUN RGB-D [128]) yielded sparse or metrically inconsistent floor reconstructions under both classical and modern multi-view fusion [89, 123, 148]. HM3DSem, the semantic subset of the HM3D dataset [111,158], has per-room mesh annotations but lacks per-room mesh seg- ments and requires a large overhead of processing meshes. SUN RGB-D [128] and SceneNN [58] add limited diversity for the additional storage overhead, but remain viable future options. S8.2 Implementation Details Hyperparameters.All single-model runs use seed42; LaMa-Ensemble uses seeds 41â44. Because the VGG perceptual loss and ResNet perceptual loss used in 46Bhattacharjee et al. F obs F â F â obs DiffusionFlowFM+XAttn LaMa Fig. S11: Supplementary failure cases.Three additional test scenes exhibiting the failure patterns identified in Sec. 6: disconnected floor islands (row 1), coarse-shape drift (row 2), and residual boundary artifacts (row 3). The boundary-leakage case is shown in the main paper (Fig. 7). The red overlay (F â obs ) shows GT floor absent from the observation (50â61% across cases). LaMa collapses in every case. the original LaMa assume 3-channel natural-image inputs, both are disabled (weight=0); the remaining losses (maskedâ 1 reconstruction, hinge adversarial loss, and multi-scale feature matching [60]) are channel-agnostic and retained unchanged (weightsλ rec =10,λ adv =10,λ fm =250; see Tab. S12). Training runs on4ĂA100 40 GB GPUs with distributed data-parallel, FP32, and activation checkpointing [26]; peak memory is 30â36 GB per GPU. At inference, we inte- grate the learned ODEs used in Flow Matching models fromt=0tot=1us- ing the second-order Heun solver [62]. We use DDIM [127] with50steps and classifier-free guidance (CFG) at scales=2.0[53] for the diffusion model. CFG at scales=2.0and per-step evidence clamping are applied for all models with dropout rate0.1. All hyperparameters follow standard literature and implemen- tations [53,77,145]. Runtime and NFE summary.Deterministic baselines require NFE=1; LaMa- Ensemble NFE=4(one forward pass per seed). Diffusion and Flow Matching use 50 solver steps per sample, while FM+XAttn uses 25 solver steps per sample. This yields NFE=200for Diffusion and Flow Matching (50ĂK) and NFE=100 for FM+XAttn (25ĂK). Deterministic models achieve>12 FPS; stochastic generators are 20â80Ăslower. FM+XAttn is the slowest despite fewer solver steps because cross-attention adds per-step overhead. We observed no significant difference in results at inference due to this change. All models were selected by peak validation micro-IoU; inference is fully deterministic given fixed noise seeds. Generative Single-View Floormap Completion47 Table S12: Training configuration and inference runtime. Top:Hyperparam- eters shared across all learned models; LaMa-specific loss weights listed separately. Bottom:Per-observation latency and throughput on a single A100 (256Ă256inputs). Training hyperparameters OptimizerAdamW [84] Learning rate1Ă10 â4 Weight decay1Ă10 â2 LR scheduleCosine annealing [83] (no restarts) Effective batch size64(16/GPUĂ4GPUs) Training steps300,000 PrecisionFP32 Gradient accumulation [9] 1 step BCE positive weight1.0(floor and non-floor equally weighted) CFG dropoutp drop 0.1(stochastic models only) LaMa / LaMa-Ensemble additional λ rec /λ adv /λ fm 10.0/10.0/250.0 Seeds (LaMa-Ensemble)41, 42, 43, 44 Inference runtime MethodSampler NFE Latency (ms) FPS U-NetForward 181 12.3 PConv-UNet Forward 16814.7 LaMaForward 163 15.9 LaMa-Ens. Forward 4136 7.4 DiffusionDDIM2001745 0.6 Flow Match. Heun2003377 0.3 FM+XAttn Heun1006661 0.2 S9 Ethical Considerations Ethical use.This work targets assistive robotics, accessibility mapping, and au- tonomous indoor navigation. The binary BEV representation abstracts away visual appearance, so the model neither processes nor generates identifiable im- agery. Inferring layout beyond the visible region could facilitate unauthorized mapping of private spaces. We recommend red-teaming, access-control policies, and audit logging at deployment. Privacy and data provenance.All six source datasets were collected with in- formed consent under their respective institutional review processes. Researchers must obtain source data under the original licenses [7, 24, 34, 36, 147, 160]. The derived BEV maps contain no personally identifiable information. Dataset Bias and environmental impact.The source corpora are predominantly North American and European interiors; models may under-perform on under- represented building typologies. Training the seven learned models required ap- proximately 280 A100-GPU hours on a shared institutional cluster powered in part by renewable energy; per-model runtime is reported in Tab. S12.