Paper deep dive
MvFFN: Multi-view Floor plan Feed-forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction
Yuguang Li, Yichuan Deng, Zixuan Liu, Ivaylo Boyadzhiev, Xiuyuan Wu, Xinyu Zou, Linda Shapiro, Alex Colburn
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/14/2026, 9:46:59 AM
Summary
The paper introduces MvFFN, a feed-forward multi-view transformer for reconstructing floor-plan layouts and camera poses from sparse, unposed, wide-baseline RGB panoramas. The core innovation is Multi-view Ordered Wall Instance Segmentation (MOIS), which predicts local wall instances and connectivity rather than global layouts. MvFFN jointly predicts coarse poses, dense geometry, and semantic wall assignments, which are then assembled via a Layout Chaining (LC) post-processing pipeline. The authors also introduce the ZInD-CrossView dataset to support this task.
Entities (8)
Relation Signals (6)
ZInD-CrossView → isaugmentationof → ZInD
confidence 95% · ZInD-CrossView dataset, which augments ZInD with globally unique wall instances
MvFFN → uses → MOIS
confidence 95% · Building on MOIS, we propose Multi-view floor-plan Feed-Forward Network (MvFFN)
MvFFN → uses → Layout Chaining
confidence 95% · A downstream Layout Chaining (LC) post processing pipeline then assembles these predictions
MvFFN → isinspiredby → VGGT
confidence 90% · MvFFN uses a multi-view geometric encoder, trained from scratch and inspired by VGGT
BADGR → demonstrates → Complementarity of semantic and geometry
confidence 85% · BADGR Li et al. (2025) shows that semantic layout correspondences and geometry are complementary
CovisPose → provides → Camera rotation priors
confidence 85% · we introduce two inductive biases based on reliable 2-view CovisPose... Camera rotation priors.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Reconstructing accurate camera poses and floor-plan layouts from sparse, unposed, wide-baseline RGB panoramas remains a challenging open problem. Direct multi-room floor-plan prediction is especially difficult in this setting: pose, geometry, semantics, and global layout structure are tightly coupled, yet a complete floor-plan is hard to learn as a single global output. Our key insight is to keep the model's predictions strictly local. Multi-view Ordered Wall Instance Segmentation (MOIS) is designed around this concept. Each panorama column predicts only the wall it observes plus the next two adjacent walls in clockwise order. The global wall partition, room layouts, and shared coordinates emerge from deterministic aggregation of overlapping local predictions. Building on MOIS, we propose Multi-view floor-plan Feed-Forward Network (MvFFN), which jointly predicts coarse camera poses, dense global geometry, and per-wall semantic outputs. A downstream Layout Chaining (LC) post processing pipeline then assembles these predictions into multi-room floor-plans with shared global coordinates, consistent wall identities, and connected room structure. The dense MOIS predictions recover cross-view wall identity and local within-room connectivity more accurately than composed modular baselines, and the assembled layouts substantially outperform prior methods on wall and junction level accuracy. We also introduce the ZInD-CrossView dataset, which augments ZInD with globally unique wall instances and cross-view segmentation labels for this task.
Tags
Links
- Source: https://arxiv.org/abs/2610.14315v1
- Canonical: https://arxiv.org/abs/2610.14315v1
Trouble viewing inline? Open PDF directly →
Full Text
92,990 characters extracted from source content.
Expand or collapse full text
MvFFN: Multi-view Floor plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction Yuguang Li †thanks: Equal contribution. 🖂 Corresponding author: ylee3@cs.washington.edu. Affiliation: University of Washington Affiliation: Yichuan Deng11footnotemark: 1 Affiliation: University of Washington Zixuan Liu11footnotemark: 1 Affiliation: University of Washington Ivaylo Boyadzhiev Affiliation: Independent Researcher Xiuyuan Wu Affiliation: University of Washington Xinyu Zou Affiliation: University of Washington Linda Shapiro Affiliation: University of Washington Alex Colburn Affiliation: University of Washington Abstract Reconstructing accurate camera poses and floor plan layouts from sparse, unposed, wide-baseline RGB panoramas remains a challenging open problem. Direct multi-room floor plan prediction is especially difficult in this setting: pose, geometry, semantics, and global layout structure are tightly coupled, yet a complete floor plan is hard to learn as a single global output. We address this challenge with Multi-view Ordered Wall Instance Segmentation (MOIS), a new formulation that replaces direct global wall-sequence prediction with local cross-view wall instances and within-room wall connectivity. Building on this formulation, we propose MvFFN, a feed-forward multi-view transformer that jointly predicts coarse camera poses, dense global geometry, and wall-instance-level semantic outputs with globally consistent identities and within-room connectivity. A downstream Layout Chaining (LC) post-processing pipeline then assembles these predictions into multi-room floor plans with shared global coordinates, consistent wall identities, and connected room structure. The resulting pipeline substantially outperforms prior methods on wall- and junction-level accuracy. Moreover, the predicted wall instances enable fully automatic downstream layout refinement, with planar bundle adjustment improving a majority of scenes and yielding clear gains on the subset that benefits. We also introduce ZInD-CrossView, a re-annotated dataset with globally unique wall instances and cross-view segmentation labels for this task. 1 Introduction Reconstructing floor plan layouts and camera poses from panoramic images is important for applications such as augmented reality, indoor digital twins, immersive walkthrough generation, and autonomous navigation. Recovering complete multi-room layouts requires multiple views because of occlusions and limited visibility. However, existing methods typically avoid the setting of unposed, wide-baseline panoramas: they assume known camera poses Yue et al. (2023); Avetisyan et al. (2024); Liu et al. (2024), require small-baseline captures Avetisyan et al. (2024), or rely on human-in-the-loop adjustment Cruz et al. (2021); Li et al. (2025). A fully automatic approach for reconstructing layouts and poses from unposed, wide-baseline panoramas would substantially reduce capture and annotation effort and enable broader deployment at scale. A floor plan encodes wall geometry, local connectivity, and room-level organization. Recovering it from multiple views requires associating wall instances across views, inferring their connectivity despite occlusion, and separating structures that belong to different rooms. Geometry alone is often insufficient. When camera poses are unknown and fused point maps are noisy, reconstruction often suffers from duplicate walls, missing walls, and incorrect junctions, which reduce both wall and junction-level precision and recall. In contrast, images provide critical semantic cues for grouping pixels across views by wall instance, inferring local connectivity, and resolving room assignments. At the same time, directly predicting a complete multi-room floor plan as a single global output is difficult to learn. Large multi-room layouts often contain ambiguous room partitions, limited observations can obscure small structural details, and small local sequence errors can cascade into large global drift. The problem is further compounded by the scarcity of datasets that provide structured multi-room layouts together with camera-registered multi-view panoramas. This motivates the central insight of our paper: rather than predicting a global floor plan directly, we recast multi-view, multi-room layout estimation around local wall relations that are easier to supervise and more robust to noisy annotations. To this end, we study structured floor plan and camera-pose reconstruction from sparse, unposed 360∘ panoramas, producing complete layouts when view coverage is sufficient and partial layouts under occlusion. We formalize this idea as Multi-view Ordered Wall Instance Segmentation (MOIS), which replaces direct global floor plan prediction with local wall instances and within-room wall connectivity that remain consistent across views. Based on MOIS, we propose MvFFN, a feed-forward multi-view network that predicts pixel-grounded wall instances and their within-room connectivity from cross-view geometric and semantic cues, and a Layout Chaining (LC) post-processing pipeline that assembles these local predictions into multi-room floor plans with consistent wall identities in a shared global frame. MvFFN also predicts coarse camera poses and dense global geometry, while the inferred wall instances enable downstream layout bundle adjustment for further refinement. Together, these design choices make multi-view layout estimation more learnable under scarce data and noisy supervision. The contributions of this work are as follows: (1) We introduce MOIS, a new formulation for multi-view, multi-room layout estimation based on local cross-view wall instances and within-room wall connectivity. (2) We introduce MvFFN, a feed-forward multi-view network that predicts these local wall structures from cross-view geometric and semantic cues. We show gains both before and after layout assembly: dense MOIS predictions improve raw cross-view wall-slot quality, while the assembled global layouts substantially improve wall and junction accuracy over prior modular baselines across both 5-view and 8-view settings. (3) We introduce ZInD-CrossView (ZInD-XV) dataset with globally unique wall instances and cross-view segmentation annotations. 2 Related Work Layout estimation. Most layout estimation methods combine a dense geometric front-end with a back-end that vectorizes predictions into wall and corner instances and room topology. In the single-view setting, methods predict monocular 1D floor boundaries Yang et al. (2019); Wang et al. (2021); Jiang et al. (2022); Sun et al. (2021) and recover wall and corner sequences through post-processing Zou et al. (2018); Sun et al. (2019); Pintore et al. (2020), relying on the clockwise image-column order within a single panorama. This cue does not directly extend to sparse multi-view settings, where each view observes only a partial layout under occlusion. For multi-room floor-plan estimation, most methods instead assume accurate camera poses and dense point clouds, and infer structure mainly from fused geometric cues in top-down space. Under imperfect geometry, connectivity is typically recovered with heuristic optimization Chen et al. (2019); Wang et al. (2022) or learned polygon and graph decoders Liu et al. (2024); Yue et al. (2023); Chen et al. (2023); Su et al. (2023). By contrast, cross-view semantic cues remain underused for resolving wall identity and connectivity under pose uncertainty. SceneScript Avetisyan et al. (2024) shows that, given camera poses and RGB video, semantic context can improve structured layout estimation. Wide-baseline pose and geometry estimation. Wide-baseline indoor pose estimation remains difficult for traditional Structure-from-Motion Schonberger and Frahm (2016); Moulon et al. (2016); Lambert et al. (2022). Recent learning-based models improve dense 3D point map and camera pose estimation from wide baseline inputs using strong foundation model encoders Wang et al. (2024); Wang et al. (2025); Tang et al. (2025). For indoor 360∘360 panoramas, limited training data often constrains geometry estimation to 1D floor boundary representations rather than dense per-pixel 3D geometry Hutchcroft et al. (2022); Tu et al. (2024); Nejatishahidin et al. (2023). BADGR Li et al. (2025) shows that semantic layout correspondences and geometry are complementary: once pixel-to-wall correspondences and initial estimates are available from unposed wide-baseline images, poses and layouts can be jointly optimized to reduce initial errors. 3 Overview Figure 1: MvFFN reformulates floor plan reconstruction from direct global prediction to local MOIS prediction: each anchor column, shown as a vertical line in the panorama, predicts cross-view masks for the anchor wall and two clockwise adjacent walls. Colors indicate corresponding wall slots/instances across panoramas and in the assembled layout. Layout Chaining aggregates these local predictions into global wall identities, room-level wall sequences, and metric floor plan layouts, enabling downstream refinement. Large multi-room floor plans often contain ambiguous room boundaries, occluded wall connectivity, and distant structural details, making direct floor-level sequence prediction brittle to error accumulation. We therefore reconstruct layouts through dense local MOIS predictions rather than direct global floor plan prediction. Our bottom-up pipeline (Figure 1) takes N panoramas and evaluates MvFFN densely at every column of every view. Each anchor column defines one MOIS target: the cross-view assignments of the current wall and its within-room neighbors. Aggregating all anchors yields dense anchor-conditioned MOIS assignments over multiple views, predicted jointly with global geometry under pose uncertainty. MvFFN uses a multi-view geometric encoder, trained from scratch and inspired by VGGT Wang et al. (2025), adapted to panoramic 2D column tokens and dense column-wise outputs. A floor-layout semantic branch combines image and geometric features for local wall-instance reasoning. We describe the geometric encoder and floor-layout decoder in Sec. 4.1 and Sec. 4.2, respectively. Finally, the dense local MOIS predictions are assembled into global room layouts via Layout Chaining (Sec. 4.5). The predicted pixel-to-wall correspondences and initial layouts can also support downstream layout refinement, which we evaluate with layout bundle adjustment in Sec. 4.6. 4 Method We formulate multi-room layout reconstruction from sparse panoramas Zhang et al. (2014) as the recovery of global 3-DoF camera poses (2D position and 1D rotation), structured 2D floor layouts, and a local anchor-conditioned wall representation across views. Inputs. We assume a set of N straightened 360∘360 RGB panoramas Iii=1N\I_i\_i=1^N Zhang et al. (2014), where each Ii∈ℝ3×H×WI_i ^3× H× W is given in equirectangular projection. The panoramas are sparsely captured in an indoor environment and are assumed to form a single pose graph whose neighboring views have at least 20%20\% co-visibility similar to Cruz et al. (2021); Hutchcroft et al. (2022); Li et al. (2025). We further assume that relative camera heights are known across the floor. MvFFN outputs. For each panorama IiI_i, MvFFN predicts four intermediate quantities: (1) a 3-DoF camera pose Ei∈SE(2)E_i∈ SE(2); (2) dense per-column floor boundaries ℬi∈ℝ2×WB_i ^2× W; (3) dense per-column floor point maps i∈ℝ4×WP_i ^4× W and wall-line parameters Λi∈ℝ3×W _i ^3× W; and (4) for each anchor column, a dense MOIS probability field S over columns across all views, encoding assignments to the underlying global wall instance intersected by the anchor ray and its within-room neighboring wall slots. MOIS formulation. We formalize the local wall-instance and within-room connectivity predictions described above with MOIS, an anchor-conditioned ordered wall representation defined for each image column (see Fig. 1 for intuition). Let i,cC_i,c denote anchor column c in panorama IiI_i. The anchor ray determines the current room by the first wall it intersects: if it intersects a closed door, the door is assigned to the attached wall instance and the current room remains the attached room; if it passes through an open doorway, the current room is the room containing the first visible wall it intersects. The corresponding MOIS target is then a local ordered K-wall tuple i,c=(wi,c(0),wi,c(1),…,wi,c(K−1)),w_i,c= (w_i,c^(0),w_i,c^(1),…,w_i,c^(K-1) ), where wi,c(0)w_i,c^(0) is the intersected wall instance defining the current room, and wi,c(r)w_i,c^(r) denotes the r-th clockwise physically adjacent wall within that room. We convert this local tuple into a dense binary assignment tensor gtS^gt, where each entry indicates whether column (j,d)(j,d) is assigned to a given ordered wall slot of the anchor column. This dense local representation serves as the semantic target for the floor-layout semantic branch in Sec. 4.2 and is assembled into global layouts by Layout Chaining in Sec. 4.5. Final layout outputs. The primary outputs are 3-DoF camera poses Eii=1N\E_i\_i=1^N and global room layouts m\W^m\. Unlike the local anchor-conditioned MOIS tuples, each mW^m represents the final reconstructed wall geometry of room m in the global floor coordinate frame after Layout Chaining. We also maintain an auxiliary column-to-wall map ℳ:i,c→wℓM:C_i,c→ w , which associates each image column with at most one global wall instance. The room corners Vm=(vm,1,…,vm,Km)V^m=(v^m,1,…,v^m,K_m) are obtained by intersecting consecutive wall segments, where vm,k∈ℝ2v^m,k ^2 and KmK_m is the number of corners. Depending on visibility, a room layout may be closed or remain open under occlusion. 4.1 N-view Geometry Encoder Branch The geometry branch is a multi-view encoder for dense geometric prediction, including camera pose, floor-boundary depth, and point maps. Architecturally, it is similar to VGGT Wang et al. (2025), while operating on panoramic column-wise features. Each panorama IiI_i is processed by an EfficientNet-v2 backbone Tan and Le (2021), followed by a Height Compression Module (HCM) Sun et al. (2019) that produces column-patch features fiG∈ℝS×Cf_i^G ^S× C, where S=256S=256, C=1024C=1024, and each patch spans 4×H4× H pixels. Patch encodings. We add angular column embeddings (ACE) and sinusoidal camera positional encodings (CPE) Vaswani (2017) to obtain geometry tokens FGF^G. The ACE are Fourier features Tancik et al. (2020) of per-patch viewing angles. For each column-patch, the viewing angle combines: (1) an intra-frame horizontal angle θi,ch∈[0∘,360∘)θ^h_i,c∈[0 ,360 ) corresponding to the patch-center azimuth, and (2) a precomputed vanishing angle θiv∈[−45∘,45∘) _i^v∈[-45 ,45 ) for camera i Zhang et al. (2014). Incorporating vanishing-angle embeddings exploits the high precision of indoor vanishing angles Li et al. (2025); Lambert et al. (2022) to axis-align the predicted floor layout. The global floor coordinate frame is centered at the first camera and oriented by its vanishing angle. Alternating-attention transformer. We concatenate geometry tokens from all views, tG=⋃i=1NfiGt^G= _i=1^Nf_i^G, with camera tokens tiC∈ℝ1×Ct_i^C ^1× C and register tokens tiR∈ℝ4×C′t_i^R ^4× C Darcet et al. (2023), and process them with a 12-layer alternating-attention transformer Wang et al. (2025). The transformer alternates between frame-wise and global attention layers implemented with FlashAttention Dao et al. (2022), and produces updated geometry, camera, and register tokens (t^iG,t^iC,t^iR)i=1N( t_i^G, t_i^C, t_i^R)_i=1^N. Dense predictions. We use separate multi-layer perceptron (MLP) heads to predict dense floor boundaries ℬ B, 2D floor point maps P, and 2D wall-line parameters Λ . The point-map output contains both 2D floor coordinates and their 2D aleatoric uncertainties. Each supporting line is represented by a 2D orientation vector and a bias term. Floor boundaries and point maps are predicted with aleatoric uncertainty Kendall and Cipolla (2016). Following prior panorama layout work Sun et al. (2019); Hutchcroft et al. (2022); Li et al. (2025), floor boundaries and point maps are supervised using derived column-wise targets that pass through open doorways and ignore occlusions from furniture and other non-structural objects, subject to dataset-specific annotation conventions. Camera poses are predicted from per-camera register and camera tokens Wang et al. (2025) using an MLP head with adaptive LayerNorm Perez et al. (2018). Figure 2: Architecture of MvFFN with two-stage training for MOIS prediction. 4.2 Floor Layout Semantic Branch Our floor-layout semantic branch predicts the dense MOIS probability field introduced in Sec. 4. Specifically, for each anchor column, it estimates which columns across all views belong to the current wall and to its next K−1K-1 clockwise physically adjacent within-room neighbors. We use K=3K=3, which provides enough neighboring-wall context for stable connectivity recovery in Layout Chaining (Sec. 4.5), while avoiding the degradation of longer prediction windows (see Appendix E.2). The branch consists of a semantic N-view encoder conditioned on geometry features, a cross-attention transformer decoder with anchor-conditioned wall-slot queries, and a segmentation head. Semantic N-view encoder. The semantic encoder follows the geometry branch architecture from Sec. 4.1. Each panorama is processed by an EfficientNet-v2 backbone and HCM to produce column-patch features fiS∈ℝS×Cf_i^S ^S× C. These features are fused with the frozen geometry tokens t^iG t_i^G through a SwiGLU module, and the fused features are then processed by a four-layer semantic alternating-attention transformer, producing semantic column tokens t^iS t_i^S. Anchor-conditioned wall-slot queries. We adopt a two-level query design with M×KM× K learned query tokens, where M is the number of sampled anchor columns and K is the number of ordered MOIS wall slots. Each query group corresponds to one anchor column and contains K ordered query tokens associated with the wall slots r∈0,…,K−1r∈\0,…,K-1\. Given sampled anchor columns (im,cm)m=1M\(i_m,c_m)\_m=1^M, we initialize query-group features using the angular encoding of the anchor column together with a camera identity encoding. Transformer decoder. The decoder has four stacked layers. Within each anchor group m, self-attention operates over the K ordered wall-slot queries, and cross-attention is applied to the semantic column tokens from all views. The decoder outputs updated query features t^m,rq t_m,r^q, one for each anchor m and wall slot r. MOIS probability prediction. Let t~m,rq∈ℝC′ t_m,r^q ^C and t~j,dI∈ℝC′ t_j,d^I ^C denote the projected query and semantic features for anchor-slot (m,r)(m,r) and column (j,d)(j,d), respectively. We compute the dense MOIS probability field as ^m,j,d,r=σ(⟨t~m,rq,t~j,dI⟩), S_m,j,d,r=σ\! ( t_m,r^q, t_j,d^I ), (1) where ^m,j,d,r S_m,j,d,r is the probability that column (j,d)(j,d) belongs to the r-th ordered wall slot of anchor m. During training, M anchor columns are randomly sampled subject to GPU memory limits. During inference, we set M=NWM=NW and predict the dense MOIS field for all anchor columns in all views in a single forward pass; equivalently a 5D tensor indexed by view and anchor column, as in Algorithm 1. 4.3 Training We train MvFFN in two stages because the geometry and semantic objectives compete under joint training. In the first stage, we train the geometry branch for dense geometric prediction, including floor boundaries, floor point maps, and camera poses, using a multi-task objective that combines geometric supervision with absolute and relative pose losses. In the second stage, we freeze the pre-trained geometry branch and train the floor-layout semantic branch from scratch under dense MOIS supervision, reusing the geometry features from stage one. The MOIS objective is factorized into a wall-presence term, which predicts whether a column belongs to any ordered wall slot of the anchor, and an ordered wall-slot classification term applied to positive assignments. In both stages, training samples are formed by drawing sparse multi-view panorama groups from a single pose graph with sufficient co-visibility. Full loss definitions, optimizer settings, and sampling schedules are provided in Appendix A. 4.4 Inductive Bias for Training Data Scarcity Training sparse multi-view models is substantially harder than training 2-view models because they require more diverse multi-view samples, while the number of scenes and unique structured training examples remains limited. In our settings, this makes naive 3-view to 5-view training unstable even when the same data is sufficient for strong 2-view geometric learning. This trend is also consistent with recent geometric models such as VGGT Wang et al. (2025) and MV-DUSt3R Tang et al. (2025), which require substantially more training data than DUSt3R Wang et al. (2024) to improve 2-view pose accuracy. To improve convergence under this regime, we introduce two inductive biases based on reliable 2-view CovisPose Hutchcroft et al. (2022) predictions. Camera rotation priors. Two-view geometric models already provide reliable camera rotations, and per-panorama vanishing points further improve indoor rotation accuracy Zhang et al. (2014); Lambert et al. (2022). We first run the 2-view CovisPose model and recover global camera rotations with a minimum spanning tree, following BADGR’s initialization strategy Li et al. (2025). These refined rotations are then injected as conditioning signals into the angular encoder (Sec. 4.1). This reduces the amount of scarce multi-view supervision spent on relearning camera rotation and lets the model focus on the harder cross-view geometric reasoning problem, while retaining near 2-view rotation accuracy without requiring substantially more N-view training data. Column-patch correspondence priors. Under limited multi-view training data, we use precomputed 2-view CovisPose correspondences during both training and inference to restrict early cross-view attention to a local window around predicted matches, rather than relearning global correspondence search from scratch. We implement this by masking the first four global self-attention layers of the alternating-attention transformer: for each column-patch token in view IaI_a, tokens in view IbI_b more than 40 columns away from the predicted match are masked out, and fully non-covisible view pairs are blocked entirely. This reduces correspondence ambiguity, leading to less noisy gradients and faster convergence. 4.5 Layout Chaining Layout Chaining (LC) is a deterministic procedure that reduces dense local MOIS predictions into global column-to-wall assignments ℳM, room-level wall sequences ℒm\L^m\, and room endpoints Vm\V^m\. Since many anchor columns provide overlapping same-wall and neighboring-wall evidence, LC first consolidates wall identity and then recovers within-room adjacency. It clusters same-view current-wall predictions ^[n,:,n,:,0] S[n,:,n,:,0] into per-view wall instances LnL_n, then merges these across views into global wall groups through a cross-view same-wall graph. Directed within-room adjacency is estimated from the neighboring MOIS slots, producing a room graph whose edges are filtered by geometric consistency and whose longest directed paths define partial room layouts ℒmL^m. Finally, wall-line parameters are fit from supporting point-map predictions grouped by ℳM, and room endpoints are recovered by intersecting consecutive walls. Algorithm details are provided in Appendix B. 4.6 Layout Bundle Adjustment The layouts produced by Layout Chaining and the predicted camera poses may still violate multi-view consistency and structural layout constraints. The predicted pixel-to-wall correspondences provide the associations needed for refinement. We investigate whether two refinement methods, planar bundle adjustment (BA) and BADGR Li et al. (2025), can use these correspondences to improve poses and wall geometry. Sec. reports the results. 5 Experiments We evaluate primarily on ZInD-XV, which extends ZInD Cruz et al. (2021) with globally indexed multi-room layouts represented as room-wise wall sequences, and per-panorama column-wise floor-boundary and wall-instance labels derived from those layouts. ZInD-XV contains 2,706 floors and 66,834 panoramas. We additionally use HM3D-XV, a smaller auxiliary benchmark derived from residential HM3DSem scenes Yadav et al. (2023), with 372 floors and 17,943 panoramas, to evaluate cross-dataset transfer to furnished, mesh-rendered indoor scenes. Dataset construction and annotation details are provided in the supplementary material. We compare against modular baselines assembled from existing pose, geometry, and layout components. The main evaluation measures final structured floor-layout accuracy after Layout Chaining; we also report pose estimation and downstream refinement with planar bundle adjustment and BADGR Li et al. (2025). To isolate the learned representation before post-processing, we additionally evaluate raw MOIS prediction quality in the supplementary material. Baseline implementation details, HM3D-XV transfer results, and additional ablations are also provided in the supplementary material. 5.1 Pose Estimation Evaluation on ZInD-XV We compare MvFFN with CovisPose+Greedy Spanning Tree (GST) Nejatishahidin et al. (2023); Li et al. (2025) and VGGT Wang et al. (2025) on ZInD-XV, adapting VGGT to panoramas via perspective decomposition. We exclude GraphCovis Nejatishahidin et al. (2023) and PanoPose Tu et al. (2024) because public code is unavailable. For each view-count setting, we evaluate on 1,000 N-view test tuples using the pose-graph alignment protocol from prior work Nejatishahidin et al. (2023); tuple construction and implementation details are provided in the supplementary. Although pose estimation is not the primary target of MvFFN, Table 1 shows that it is comparable to CovisPose+GST: CovisPose+GST attains lower median errors, while MvFFN reduces mean error, variance, and hard-case failures. VGGT performs substantially worse in this sparse wide-baseline panorama setting. Table 1: Camera pose evaluation on ZInD-XV dataset. Methods 5v Pos (cm)↓ 5v Rot (∘)↓ 8v Pos (cm)↓ 8v Rot (∘)↓ Mn Std Med Mn Std Med Mn Std Med Mn Std Med CovisPose+GST 12.6 23.3 9.3 1.3 0.9 0.7 17.3 37.3 11.8 1.9 1.6 0.8 VGGT 30.7 39.4 13.3 11.1 16.0 4.0 53.4 63.1 27.8 13.3 17.7 4.9 MvFFN 12.5 10.8 10.2 1.2 0.6 1.1 17.0 15.6 13.8 1.3 0.8 1.2 5.2 Floor Layout Estimation Evaluation on ZInD-XV We compare MvFFN with modular floor-layout pipelines because no prior method directly predicts structured multi-room layouts from sparse, unposed N-view panoramas. As summarized in Table 2, each baseline composes pose estimation, geometric input construction, and layout vectorization: CovisPose+GST (CP) or ground-truth poses provide the pose input, LGT-Net (LGT) provides panorama-level boundary cues, and the final layout is produced by either heuristic LGT post-processing (LGT-P) or learning-based top-down vectorization with RoomFormer (RF) Yue et al. (2023) and PolyRoom (PR) Liu et al. (2024). We exclude VGGT from layout evaluation because its pose accuracy is substantially worse than CovisPose; implementation and metric details are provided in the supplementary material. Table 2: Floor-layout evaluation on ZInD-XV. Distance is in cm. Geom. denotes the intermediate geometric input and Layout denotes the vectorization backend. GT-pose denotes a control group. Modular Pipeline 5-view wall 5-view junction 8-view wall 8-view junction pose geom. layouts Prec.↑ Rec.↑ F1 ↑ Dist ↓ Prec.↑ Rec.↑ F1 ↑ Dist ↓ Prec.↑ Rec.↑ F1 ↑ Dist ↓ Prec.↑ Rec.↑ F1 ↑ Dist ↓ MvFFN 88.5 87.7 88.1 13.9 79.2 68.2 73.3 13.8 89.4 87.5 88.4 19.0 81.0 69.8 75.0 19.1 CP RF 31.9 44.2 37.1 84.9 15.4 20.9 17.7 137.8 35.0 42.4 38.3 79.0 15.7 18.6 17.0 136.8 CP PR 33.6 29.3 31.3 86.6 17.4 14.9 16.1 129.7 31.8 25.0 28.0 86.2 13.8 10.5 11.9 133.4 CP LGT LGT-P 75.6 71.4 73.4 21.3 60.0 54.8 57.3 21.3 76.3 72.7 74.5 24.4 61.7 56.7 59.1 24.4 GT LGT LGT-P 78.8 70.5 74.4 19.2 62.4 54.3 58.1 19.7 80.5 70.8 75.3 22.7 65.1 55.5 59.9 22.7 We evaluate wall and junction precision, recall, and F1F1 with instance-level matching, and report matched-wall localization error; details are in the supplementary material. These structure-aware metrics are better suited to our setting than region IoUIoU, since high-quality floor plan reconstruction requires accurate wall instances, junctions, connectivity, and geometric alignment, and MvFFN may produce either closed polygons or open wall sequences under limited view coverage Cruz et al. (2021). As shown in Table 2, MvFFN substantially outperforms all modular baselines across both 5-view and 8-view settings, with the largest gains over learned top-down vectorization backends and consistent improvements over the strongest LGT-P baseline. Notably, MvFFN also outperforms LGT-P with ground-truth poses, indicating that the gains are not explained merely by better pose estimation. Instead, the wall and junction improvements show that cross-view MOIS prediction and Layout Chaining recover wall identities, within-room connectivity, and global layout structure more reliably than modular geometry-first pipelines. 5.3 Global Refinement We evaluate whether MvFFN’s predicted wall instances and pixel-to-wall correspondences support automatic downstream refinement. Table 3 reports matched-wall distance before and after planar bundle adjustment (BA) and BADGR Li et al. (2025), both over all samples and over the subset improved by each method. Planar bundle adjustment improves a majority of samples in both 5-view and 8-view settings, with clear gains on the improved subset. Overall gains are modest, and BADGR is more sensitive to initialization because it assumes complete initial room layouts and clean correspondences. Details are provided in the supplementary material. Table 3: Matched-wall distance before and after refinement on ZInD-XV. Distances are in cm. View Method All samples Improved samples Mean dist Med. dist Improved (%) Mean dist Med. dist before after before after before after before after 5-view MvFFN+BA 13.9 13.7 6.9 6.7 68.0% 14.8 12.3 7.5 5.9 MvFFN+BADGR 13.9 17.7 6.9 8.1 45.9% 16.2 14.1 8.2 6.2 8-view MvFFN+BA 19.0 18.8 8.8 8.8 66.0% 20.6 17.8 9.7 8.2 MvFFN+BADGR 19.0 21.0 10.1 10.4 48.5% 22.9 20.0 8.8 8.8 6 Ablation Studies We evaluate three complementary ablations. The main paper studies architecture, testing whether the two-branch, two-stage MvFFN design is necessary for stable MOIS prediction. Appendix E.1 studies inductive bias, evaluating rotation and correspondence priors across input-view counts, and Appendix E.2 studies prediction window, where longer MOIS wall sequences become harder to learn. Architecture. We ablate two core design choices in MvFFN: a dedicated semantic branch for MOIS prediction and the two-stage geometry-to-semantics training strategy. Table 4 compares four variants: a shared-encoder model, a two-branch model without geometry-to-semantic fusion, a full architecture trained in one stage, and the full two-stage MvFFN. The results show that both geometry-conditioned semantic reasoning and staged training are critical: removing geometry-to-semantic fusion sharply reduces recall and F1F1, while shared features or single-stage training underperform and become unstable, indicating tension between dense geometry objectives and the semantic MOIS objective. The full two-stage design gives the best stability and final layout accuracy. Table 4: Ablation study of MvFFN on ZInD-XV, comparing encoder design, geometry fusion, and staged training. – denotes pose metrics identical to MvFFN. Method Geom Enc Sem Enc Geom Fusion Train Stages 5v pos (cm)↓ 5v rot (∘)↓ 5v wall 5v junction Mn Std Med Mn Std Med Prec↑ Rec↑ F1↑F1 Dist↓ Prec↑ Rec↑ F1↑F1 Dist↓ Shared-Enc ✓ ✗ ✗ 2 35.8 50.1 26.4 2.1 1.3 2.0 86.4 69.8 77.2 31.2 72.9 50.4 59.6 39.9 No-Fusion ✓ ✓ ✗ 2 – – – – – – 92.6 22.7 36.5 28.0 84.3 14.4 24.6 21.4 1-stage ✓ ✓ ✓ 1 46.8 50.5 30.1 2.5 1.5 1.9 87.7 71.9 79.0 33.9 75.8 53.1 62.5 44.1 MvFFN ✓ ✓ ✓ 2 12.5 10.8 10.2 1.2 0.6 1.1 88.5 87.7 88.1 13.9 79.2 68.2 73.3 13.8 7 Qualitative Results Qualitative results in Fig. 3 show representative 8-view reconstructions, with additional examples provided in Appendix J. Across simple and complex multi-room scenes, MvFFN produces more complete and globally consistent layouts than the strongest modular baseline, LGT-Net with CovisPose. The predicted wall instances are color-coded to show cross-view consistency. Compared with the baseline, MvFFN better preserves room shape, avoids duplicate or fragmented walls, and recovers neighboring structure visible through doorways under partial occlusion. Figure 3: Qualitative comparisons on ZInD-XV test set. We show the strongest modular baseline, LGT-Net with CovisPose. 8 Conclusion We introduced MvFFN, which reconstructs structured floor plans and camera poses from sparse, unposed 360∘360 panoramas by predicting cross-view wall instances and within-room connectivity before global assembly. Experiments show clear gains over modular baselines, demonstrate automatic downstream refinement from predicted wall correspondences, and introduce ZInD-XV for future structured multi-view layout research. References Avetisyan et al. (2024) A. Avetisyan, C. Xie, H. Howard-Jenkins, T. Yang, S. Aroudj, S. Patra, F. Zhang, D. Frost, L. Holland, C. Orme, et al. Scenescript: reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision, p. 247–263. Cited by: §1, §2. Bridson (2007) R. Bridson Fast Poisson disk sampling in arbitrary dimensions. In ACM SIGGRAPH 2007 Sketches, p. 22–es. Cited by: §F.2. Chen et al. (2023) J. Chen, R. Deng, and Y. Furukawa PolyDiffuse: polygonal shape reconstruction via guided set diffusion models. arXiv preprint arXiv:2306.01461. Cited by: §2. Chen et al. (2019) J. Chen, C. Liu, J. Wu, and Y. Furukawa Floor-sp: inverse cad for floorplans by sequential room-wise shortest path. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2661–2670. Cited by: §2. Cheng et al. (2021) B. Cheng, A. Schwing, and A. Kirillov Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems 34, p. 17864–17875. Cited by: §G.2. Cruz et al. (2021) S. Cruz, W. Hutchcroft, Y. Li, N. Khosravan, I. Boyadzhiev, and S. B. Kang Zillow indoor dataset: annotated floor plans with 360deg panoramas and 3d room layouts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 2133–2143. Cited by: Appendix K, §F.1, §G.1, §1, §4, §5.2, §5. Dao et al. (2022) T. Dao, D. Fu, S. Ermon, A. Rudra, and C. Ré Flashattention: fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems 35, p. 16344–16359. Cited by: §4.1. Darcet et al. (2023) T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski Vision transformers need registers. arXiv preprint arXiv:2309.16588. Cited by: §4.1. Hutchcroft et al. (2022) W. Hutchcroft, Y. Li, I. Boyadzhiev, Z. Wan, H. Wang, and S. B. Kang CoVisPose: co-visibility pose transformer for wide-baseline relative pose estimation in 360-degree indoor panoramas. In European Conference on Computer Vision, p. 615–633. Cited by: Appendix A, Appendix C, §F.1, §2, §4.1, §4.4, §4. Jiang et al. (2022) Z. Jiang, Z. Xiang, J. Xu, and M. Zhao Lgt-net: indoor panoramic room layout estimation with geometry-aware transformer network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 1654–1663. Cited by: Appendix C, §G.2, §2. Kendall and Cipolla (2016) A. Kendall and R. Cipolla Modelling uncertainty in deep learning for camera relocalization. In 2016 IEEE international conference on Robotics and Automation (ICRA), p. 4762–4769. Cited by: Appendix A, §4.1. Lambert et al. (2022) J. Lambert, Y. Li, I. Boyadzhiev, L. Wixson, M. Narayana, W. Hutchcroft, J. Hays, F. Dellaert, and S. B. Kang Salve: semantic alignment verification for floorplan reconstruction from sparse panoramas. In European Conference on Computer Vision, p. 647–664. Cited by: §2, §4.1, §4.4. Li et al. (2025) Y. Li, I. Boyadzhiev, Z. Liu, L. Shapiro, and A. Colburn BADGR: bundle adjustment diffusion conditioned by gradients for wide-baseline floor plan reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 16785–16795. Cited by: §F.1, §F.2, §G.1, §G.1, §G.4, §1, §2, §4.1, §4.1, §4.4, §4.6, §4, §5.1, §5.3, §5. Liu et al. (2024) Y. Liu, L. Zhu, X. Ma, H. Ye, X. Gao, X. Zheng, and S. Shen PolyRoom: room-aware transformer for floorplan reconstruction. arXiv preprint arXiv:2407.10439. Cited by: §G.2, §G.2, §1, §2, §5.2. Moulon et al. (2016) P. Moulon, P. Monasse, R. Perrot, and R. Marlet OpenMVG: open multiple view geometry. In International Workshop on Reproducible Research in Pattern Recognition, p. 60–74. Cited by: §2. Nejatishahidin et al. (2023) N. Nejatishahidin, W. Hutchcroft, M. Narayana, I. Boyadzhiev, Y. Li, N. Khosravan, J. Košecká, and S. B. Kang Graph-covis: gnn-based multi-view panorama global pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 6459–6468. Cited by: §G.1, §G.1, §2, §5.1. Perez et al. (2018) E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville Film: visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32. Cited by: §4.1. Pintore et al. (2020) G. Pintore, M. Agus, and E. Gobbetti AtlantaNet: inferring the 3d indoor layout from a single 360-degree image beyond the manhattan world assumption. In European Conference on Computer Vision, p. 432–448. Cited by: §2. Ramakrishnan et al. (2021) S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238. Cited by: §F.2. Schonberger and Frahm (2016) J. L. Schonberger and J. Frahm Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 4104–4113. Cited by: §2. Su et al. (2023) J. Su, K. Tung, C. Peng, P. Wonka, and H. J. Chu Slibo-net: floorplan reconstruction via slicing box representation with local geometry regularization. Advances in Neural Information Processing Systems 36, p. 48781–48792. Cited by: §2. Sun et al. (2019) C. Sun, C. Hsiao, M. Sun, and H. Chen Horizonnet: learning room layout with 1d representation and pano stretch data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 1047–1056. Cited by: §2, §4.1, §4.1. Sun et al. (2021) C. Sun, M. Sun, and H. Chen Hohonet: 360 indoor holistic understanding with latent horizontal features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2573–2582. Cited by: §2. Tan and Le (2021) M. Tan and Q. Le Efficientnetv2: smaller models and faster training. In International conference on machine learning, p. 10096–10106. Cited by: §4.1. Tancik et al. (2020) M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems 33, p. 7537–7547. Cited by: §4.1. Tang et al. (2025) Z. Tang, Y. Fan, D. Wang, H. Xu, R. Ranjan, A. Schwing, and Z. Yan Mv-dust3r+: single-stage scene reconstruction from sparse views in 2 seconds. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 5283–5293. Cited by: §2, §4.4. Tu et al. (2024) D. Tu, H. Cui, X. Zheng, and S. Shen PanoPose: self-supervised relative pose estimation for panoramic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20009–20018. Cited by: §G.1, §2, §5.1. Vaswani (2017) A. Vaswani Attention is all you need. Advances in Neural Information Processing Systems. Cited by: §4.1. Wang et al. (2021) F. Wang, Y. Yeh, M. Sun, W. Chiu, and Y. Tsai Led2-net: monocular 360deg layout estimation via differentiable depth rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12956–12965. Cited by: §2. Wang et al. (2022) H. Wang, W. Hutchcroft, Y. Li, Z. Wan, I. Boyadzhiev, Y. Tian, and S. B. Kang Psmnet: position-aware stereo merging network for room layout estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8616–8625. Cited by: §2. Wang et al. (2025) J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny Vggt: visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 5294–5306. Cited by: §G.1, §2, §3, §4.1, §4.1, §4.1, §4.4, §5.1. Wang et al. (2024) S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud Dust3r: geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20697–20709. Cited by: Appendix A, §2, §4.4. Yadav et al. (2023) K. Yadav, R. Ramrakhya, S. K. Ramakrishnan, T. Gervet, J. Turner, A. Gokaslan, N. Maestre, A. X. Chang, D. Batra, M. Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 4927–4936. Cited by: Appendix K, Appendix D, §F.2, §5. Yang et al. (2019) S. Yang, F. Wang, C. Peng, P. Wonka, M. Sun, and H. Chu Dula-net: a dual-projection network for estimating room layouts from a single rgb panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3363–3372. Cited by: §2. Yue et al. (2023) Y. Yue, T. Kontogianni, K. Schindler, and F. Engelmann Connecting the dots: floorplan reconstruction using two-level queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 845–854. Cited by: §G.2, §G.2, §1, §2, §5.2. Zhang et al. (2014) Y. Zhang, S. Song, P. Tan, and J. Xiao Panocontext: a whole-room 3d context model for panoramic scene understanding. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, p. 668–686. Cited by: §4.1, §4.4, §4, §4. Zhou (2023) L. Zhou Efficient second-order plane adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 13113–13121. Cited by: §G.4. Zou et al. (2018) C. Zou, A. Colburn, Q. Shan, and D. Hoiem Layoutnet: reconstructing the 3d room layout from a single rgb image. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 2051–2059. Cited by: §2. Appendix Appendix A Training Details We train MvFFN in two stages. N-view geometry pre-training. We first train the geometry branch with the multi-task loss ℒG=ℒ+ℒℬ+ℒrel+ℒabs.L^G=L_P+L_B+L_rel+L_abs. (2) Here, ℒL_P and ℒℬL_B denote the point-map and floor boundary losses. Following DUSt3R Wang et al. [2024], both use aleatoric uncertainty weighting Kendall and Cipolla [2016] with predicted uncertainties Σi _i^P and Σiℬ _i^B. The pose losses ℒrelL_rel and ℒabsL_abs follow CovisPose Hutchcroft et al. [2022]; valid pairs for ℒrelL_rel have minimum co-visibility 0.01. Specifically, ℒ _P =∑i=1N‖Σi(i−^i)‖−αlogΣi, = _i=1^N _i^P(P_i- P_i) -α _i^P, ℒℬ _B =∑i=1N‖Σiℬ(ℬi−ℬ^i)‖−αlogΣiℬ, = _i=1^N _i^B(B_i- B_i) -α _i^B, ℒabs _abs =∑i=1N(∥ti−t^i∥+βr∥ri−r^i∥). = _i=1^N ( t_i- t_i + _r r_i- r_i ). and ℒrelL_rel is defined analogously on relative camera translations and rotations for valid image pairs. Here, rir_i, tit_i and r^i r_i, t^i t_i denote the ground truth and predicted camera translation and normalized rotation vector for camera i, and βr=0.1 _r=0.1. We optimize Eq. (2) with AdamW for 80k iterations, a cosine learning-rate schedule with peak 10−410^-4, and 4k warmup iterations. Each batch samples one scene and 2, 3, 5, or 8 images per example such that the views form a single pose graph with co-visibility greater than 20%20\%. A linear view-count scheduler shifts from 2 and 3-view batches early in training to 5 and 8-view batches later. We use random horizontal rotation augmentation and clip gradients to norm 1.0. Layout semantic training. We train the floor-layout semantic branch from scratch while freezing the geometry branch and reusing its features. We use the same data-sampling schedule and a similar learning-rate schedule, with maximum learning rate 3.5×10−43.5× 10^-4. The semantic branch uses dense MOIS supervision with a factorized objective: wall-presence prediction for each anchor-column pair, and ordered wall-slot classification on positives. Let m∈1,…,Mm∈\1,…,M\ index sampled anchor columns, with anchor m corresponding to (im,cm)(i_m,c_m). For anchor m and column (j,d)(j,d), the decoder produces dense MOIS probabilities ^m,j,d,r=σ(zm,j,d,rcls) S_m,j,d,r=σ(z^cls_m,j,d,r), where r∈0,…,K−1r∈\0,…,K-1\ indexes the ordered wall slot relative to the anchor. The binary supervision is m,j,d,rgt∈0,1S^gt_m,j,d,r∈\0,1\. We aggregate wall-slot logits into a presence logit zm,j,dpres=log∑r=0K−1exp(zm,j,d,rcls)z_m,j,d^pres= _r=0^K-1 (z_m,j,d,r^cls), with target ym,j,dpres∈0,1y_m,j,d^pres∈\0,1\ indicating whether (j,d)(j,d) is assigned to any ordered wall slot of anchor m. The wall-presence and ordered-slot classification losses are ℒpres _pres =1|Ωvalid|∑(m,j,d)∈ΩvalidBinaryCrossEntropy(zm,j,dpres,ym,j,dpres), = 1| _valid| _(m,j,d)∈ _validBinaryCrossEntropy\! (z_m,j,d^pres,\,y_m,j,d^pres ), ℒcls _cls =1|Ωpos|∑(m,j,d)∈ΩposCrossEntropy(zm,j,d,:cls,ym,j,dcls), = 1| _pos| _(m,j,d)∈ _posCrossEntropy\! (z_m,j,d,:^cls,\,y_m,j,d^cls ), ℒMOIS _MOIS =ℒpres+λclsℒcls. =L_pres+ _clsL_cls. Here, Ωvalid _valid is the set of valid projected panorama columns after masking invalid regions; Ωpos⊆Ωvalid _pos _valid contains columns assigned to one of the K ordered slots, with label ym,j,dcls∈0,…,K−1y_m,j,d^cls∈\0,…,K-1\. Appendix B Layout Chaining Post Processing Layout Chaining (LC) decodes the dense MOIS tensor S into compact structured layout outputs: global column-to-wall assignments ℳM, room-level wall sequences ℒm\L^m\, and metric wall endpoints Vm\V^m\. The input is dense because every anchor column predicts cross-view masks for the current wall and neighboring wall slots, creating redundant local evidence for wall identity and adjacency. LC reduces this evidence in four stages: (1) same-view current-wall masks are grouped into per-view wall instances LnL_n; (2) cross-view same-wall evidence merges them into global wall instances, defining ℳM; (3) neighboring-wall slots induce directed within-room adjacency and room-level wall sequences ℒm\L^m\; and (4) predicted point maps and wall-line parameters provide metric support for wall-line fitting and endpoint recovery, yielding Vm\V^m\. Algorithm 1 MvFFN Post-processing: Layout Chaining Input: Dense MOIS field ^∈[0,1]N×W×N×W×K S∈[0,1]^N× W× N× W× K, floor point maps ^i,c\ P_i,c\, and wall-line parameters λ^i,c\ λ_i,c\ Output: Column-to-wall assignments ℳM, within-room connectivity ℒm\L^m\, and room endpoints Vm\V^m\ /* Stage 1: Per-view wall instances */ 1 for each view n do 2 Define per-view current wall segmentation: ^nview=^[n,:,n,:,0] S^view_n= S[n,:,n,:,0]; 3 Define per-view column nodes: nview=1,…,WV^view_n=\1,…,W\; 4 for column pair (c1,c2)∈nview×nview(c_1,c_2) ^view_n×V^view_n do 5 Define same-wall column edge: add (c1,c2)(c_1,c_2) to ℰnviewE^view_n if IoU(^nview[c1],^nview[c2])>τiouIoU( S^view_n[c_1], S^view_n[c_2])> _iou; 6 end for 7 Build per-view same-wall column graph: nview=(nview,ℰnview)G^view_n=(V^view_n,E^view_n); 8 Collect per-view wall instances: Ln=ConnectedComponents(nview)L_n=ConnectedComponents(G^view_n); 9 for each connected component l∈Lnl∈ L_n do 10 Select anchor column: c∗=argmaxc∈ldeg(c)c^*= _c∈ l (c); 11 Define instance mask: l∗=^nview[c∗]l^*= S^view_n[c^*]; 12 end for 13 end for /* Stage 2: Cross-view global wall instances */ 14 Define cross-view wall nodes: cross=⋃n=1NLnV_cross= _n=1^NL_n; 15 for instance pair (li,lj)∈Ln1×Ln2,n1≠n2(l_i,l_j)∈ L_n_1× L_n_2,\;n_1≠ n_2 do 16 Define cross-view same-wall instance edge: add (li,lj)(l_i,l_j) to ℰcrossE_cross if IoU(li∗,lj∗)>τiouIoU(l_i^*,l_j^*)> _iou; 17 end for 18 Build cross-view same-wall instance graph: cross=(cross,ℰcross)G_cross=(V_cross,E_cross); 19 Collect global wall instances via cross-view column-to-wall assignments: ℳ=ConnectedComponents(cross)M=ConnectedComponents(G_cross); /* Stage 3: Room connectivity and partial layouts */ 20 Define room-wall nodes: wℓ\w \, induced by ℳM; 21 Define directed room-wall edges: ℰroom=(wi,wj)∣the k→0 predictions in ^[:,:,:,:,1:K−1] support wi→wjE_room=\(w_i,w_j) k\!→\!0 predictions in S[:,:,:,:,1:K\!-\!1] support w_i→ w_j\; 22 Build directed room graph: room=(wℓ,ℰroom)G_room=(\w \,E_room); 23 Filter invalid edges by geometry: room=GeomFilter(room)G_room=GeomFilter(G_room); 24 for each weakly connected component ℋm⊆roomH^m _room do 25 Collect within-room connectivity: ℒm=LongestPath(ℋm)L^m=LongestPath(H^m); 26 end for /* Stage 4: Final wall geometry */ 27 Define global wall parameters: (b^ℓ,^ℓ)=LineFit(^i,c,ℳ)\( b , v )\=LineFit(\ P_i,c\,M); 28 for each partial room layout ℒmL^m do 29 Define room endpoints: Vm=(vm,1,…,vm,Km)=LineIntersect(ℒm,ℳ)V^m=(v^m,1,…,v^m,K_m)=LineIntersect(L^m,M); 30 end for Appendix C Intermediate MOIS Prediction Evaluation MOIS evaluation. We evaluate the intermediate MOIS representation before Layout Chaining with Pooled Ordered Wall-Slot IoU, an IoUIoU metric computed separately for each ordered MOIS wall slot. For each N-view test tuple, we sample anchor columns from the conditioning panoramas and retain anchors with valid ordered slot targets. Each anchor instantiates the cross-view slot masks defined in Sec. 4 of the main paper. For each slot, we accumulate intersections and unions between the predicted and ground-truth masks over all valid anchors, views, and panorama columns before taking the ratio. This pixel-weighted pooling measures aggregate cross-view mask quality and avoids the high variance of query-averaged mIoUmIoU for small or weakly visible wall instances. LGT-Net + CovisPose baseline. No prior method directly predicts ordered cross-view wall instances, so we construct a composed baseline from LGT-Net Jiang et al. [2022] and CovisPose Hutchcroft et al. [2022]. The goal is to approximate MOIS by first assigning each anchor column to an ordered set of local wall slots in its source view, then propagating those slots across views. Given the single-view layout predicted by LGT-Net, we convert it into a floor-plan polygon whose edges define the local slots: the edge intersected by the anchor ray is slot 1, and the next two adjacent edges in clockwise order are slots 2 and 3. CovisPose then provides cross-view column correspondences, which propagate these source-view slots to other panoramas and produce cross-view slot masks. We evaluate these masks against the same ordered wall-instance targets using Pooled Ordered Wall-Slot IoU. Table C.1: Pooled Ordered Wall-Slot IoU↑ (%) for raw MOIS predictions before Layout Chaining. Wall 1 denotes the anchor wall, and 2-3 denote the next ordered neighboring walls. 5-view inputs 8-view inputs Method wall 1 wall 2 wall 3 wall 1 wall 2 wall 3 MvFFN 81.9 71.1 57.8 75.4 63.8 47.7 LGT + CP 63.7 37.8 20.3 60.4 34.5 18.5 Table C.1 shows that MvFFN substantially improves raw MOIS prediction quality before any Layout Chaining post-processing. Wall 1 mainly reflects cross-view wall-instance recovery, while walls 2 and 3 evaluate the ordered neighboring-wall predictions that encode local within-room connectivity. The larger gains on walls 2 and 3 show that MvFFN learns not only same-wall correspondence but also adjacent-wall structure more reliably than the composed LGT-Net+CovisPose baseline. These stronger local predictions provide better inputs to Layout Chaining, supporting the downstream gains in final wall and junction accuracy. Appendix D Experiments on HM3D-XV Dataset We use HM3D-XV as an auxiliary cross-domain validation benchmark to evaluate whether MvFFN and the MOIS formulation remain effective beyond ZInD-XV. HM3D-XV is derived from residential HM3DSem scenes Yadav et al. [2023] with sufficient mesh coverage, and contains 17,943 panoramas and 372 floors from 180 scenes. Compared with ZInD-XV, which contains 2,706 floors, HM3D-XV is smaller and noisier, contains furnished mesh-rendered scenes with rendering artifacts, and has sparser column-level supervision, with 27.3% of image columns masked after annotation cleanup. Dataset construction details are provided in Sec. F.2. This makes HM3D-XV a challenging setting for both pose estimation and layout reconstruction. We therefore use ZInD-XV as an anchor dataset and finetune MvFFN with mixed ZInD-XV/HM3D-XV samples to test whether the learned geometry and MOIS representations remain useful under domain shift and noisy annotations. For evaluation, we construct N-view test splits from per-floor tuples generated by covisibility-based sampling. Each floor receives a sample count proportional to its number of panoramas, capped by the number of valid tuples; remaining samples are redistributed according to residual capacity. This yields 951 tuples for the 5-view split and 991 tuples for the 8-view split, spanning all floors with valid N-view connected components while keeping floor-level representation roughly proportional to scene size. Both pose and layout experiments use these splits. Because some comparisons rely on CovisPose-derived priors, we additionally report those results on the CovisPose-feasible subset defined in Sec. D.1. Sec. D.1 evaluates pose estimation, and Sec. D.2 evaluates floor layout reconstruction. D.1 Camera Pose Estimation We follow the same pose-evaluation protocol as the ZInD-XV experiments and report pose accuracy on a CovisPose-feasible subset of HM3D-XV. Starting from the ZInD-XV-trained model, we fine-tune the geometry branch on HM3D-XV using the same losses, view-count schedule, and dataloader setup as Sec. A, with a 60k-iteration learning-rate schedule and the 40k-iteration checkpoint used for evaluation. Training batches mix HM3D-XV and ZInD-XV examples at a 1:1 ratio to adapt to HM3D-XV while retaining broader geometric variation from ZInD-XV. Furnished, mesh-rendered HM3D-XV panoramas differ substantially from the ZInD panoramas used to train the available CovisPose checkpoint. As a result, some pairwise CovisPose estimates are unreliable. We filter pairs with rotation error above 30∘30 or translation error above 2 m, and retain only tuples whose remaining estimates form a connected GST pose graph. This removes invalid pairwise links that would otherwise dominate pose and layout metrics, retaining approximately 60% of the 5-view tuples and 65% of the 8-view tuples. We compare CovisPose+GST, VGGT, and MvFFN. For MvFFN, we use the same rotation-prior construction as in ZInD-XV, with CovisPose rotations refined by vanishing-angle correction. For the correspondence attention mask, we use ground-truth column-patch correspondences to provide a reliable inductive bias for sparse multi-view training on HM3D-XV. Table D.1: Camera pose evaluation on the CovisPose-feasible subset of HM3D-XV. Mn/Std/Med denote mean, standard deviation, and median errors. Methods 5v Pos (cm)↓ 5v Rot (∘)↓ 8v Pos (cm)↓ 8v Rot (∘)↓ Mn Std Med Mn Std Med Mn Std Med Mn Std Med VGGT 35.2 109.3 10.4 6.1 16.3 2.1 62.3 71.8 37.4 13.7 6.4 12.3 CovisPose+GST 30.1 24.7 19.3 2.2 3.3 1.1 44.0 59.7 25.9 2.9 4.4 1.4 MvFFN 26.8 21.2 21.5 1.8 0.6 1.6 32.0 23.5 26.6 1.9 0.5 1.7 As shown in Table D.1, under reliable correspondence initialization, MvFFN produces stable pose estimates on the CovisPose-feasible subset of HM3D-XV. Compared with CovisPose+GST, it achieves lower mean errors and lower variance across both 5-view and 8-view settings, while CovisPose+GST and VGGT obtain lower medians in some cases. These results suggest that the MvFFN geometry branch, under this setting, can support stable multi-view pose estimation in the furnished HM3D-XV domain. D.2 Layout Estimation For layout estimation, our goal is to test whether the learned MOIS representation and layout decoder can support multi-room reconstruction in the furnished, mesh-rendered HM3D-XV domain. Following the pose evaluation, we use the same prior setup for MvFFN as in Sec. D.1. Starting from the HM3D-XV tuned geometry checkpoint, we retrain only the floor-layout semantic branch with mixed ZInD-XV and HM3D-XV samples, using a 2:1 ZInD-XV: HM3D-XV sampling ratio and otherwise keeping the training settings the same as Sec. A. We compare against LGT-Net based modular baselines using either CovisPose or ground truth camera poses. As in Sec. 5.2 of the main paper, the modular pipeline separates panorama-level boundary prediction from layout vectorization: LGT-Net (LGT) provides floor boundary predictions, and LGT-Net post-processing (LGT-P) converts them into vectorized layouts. RoomFormer and PolyRoom are omitted because they are substantially weaker on the main ZInD-XV layout benchmark. Table D.2: Floor-layout evaluation on HM3D-XV. Distance is in cm. Geom. denotes the intermediate geometric input and Layout denotes the vectorization backend. GT-pose denotes a control group. pose pts layouts 5-view wall 5-view junction 8-view wall 8-view junction prec recall f1 dist prec recall f1 dist prec recall f1 dist prec recall f1 dist CP LGT LGT 57.0 62.6 59.7 51.8 18.2 33.3 23.5 53.1 59.4 60.2 59.8 52.1 19.0 30.5 23.4 53.8 GT LGT LGT 69.1 66.8 67.9 36.9 25.5 44.3 32.4 35.7 69.7 64.6 67.1 39.9 25.5 38.9 30.8 38.6 MvFFN 77.8 69.6 73.5 22.1 55.9 38.8 45.8 24.1 75.0 71.6 73.3 28.9 52.9 35.0 42.1 29.3 As shown in Table D.2, under the reliable prior setup, MvFFN substantially improves layout accuracy on HM3D-XV despite the smaller and noisier training set. It achieves much higher wall precision and F1F1 than the LGT-Net based baselines, and reduces matched wall distance in both 5-view and 8-view settings. The gains are especially large at the junction level, where the modular baselines remain weak even with ground truth poses. This suggests that MOIS remains an effective supervision target under sparse column-level annotations, and helps MvFFN recover wall connectivity and junction structure more reliably than geometry-first modular pipelines. Appendix E More Ablation Studies The main paper reports the architecture ablation for the two-branch, two-stage MvFFN design. Here we provide two additional ablations: Sec. E.1 studies the rotation and correspondence priors used for sparse multi-view training, and Sec. E.2 evaluates larger MOIS prediction windows to test the effect of local prediction length. E.1 Inductive Bias Figure E.1: Effect of inductive bias across input views. Box plots show translation (left) and rotation (right) errors under different inductive-bias variants. Dashed lines and diamond markers indicate mean errors. Stronger inductive bias improves translation accuracy and suppresses extreme rotation outliers. We evaluate MvFFN across varying numbers of input panoramas, using three levels of inductive bias described in Sec. 4.4 of the main paper: no inductive bias, rotation priors only, and rotation priors with correspondence-based attention masks. As shown in Fig. E.1, stronger inductive bias has its clearest effect on translation accuracy: the full MvFFN maintains substantially lower translation error across all input-view settings, while the weaker-bias variants degrade much more rapidly as the number of views increases. For rotation, the main benefit is robustness rather than uniformly lower central statistics. In some settings, the no-bias variant can appear slightly better in the median, but this occurs only near the center of the distribution, where all methods already achieve low angular error. Its mean is nonetheless much higher because the error distribution develops a heavy tail: in the 8-view setting, the no-bias variant reaches 17.5∘17.5 at p90 and 27.8∘27.8 at p95, whereas the full MvFFN remains at 2.5∘2.5 at p95. Thus, the proposed inductive bias mainly suppresses large rotation failures and produces much more stable estimates in hard cases. Since all variants already use vanishing-angle priors, this gain is not explained by rotation snapping alone, but by the additional regularization introduced by the multi-view inductive bias. E.2 MOIS Prediction Window We train two additional MvFFN variants with larger MOIS prediction windows of 5 and 7 ordered walls, and report the results in Table E.1. For a controlled comparison, Layout Chaining is kept unchanged and uses only the first three predicted wall slots for all variants, so the ablation isolates the effect of training MvFFN to predict longer MOIS windows. Expanding the prediction window leads to a consistent drop in layout accuracy, with 5-view wall F1F1 decreasing from 88.1 to 85.6 and 84.4, and 8-view junction F1F1 decreasing from 75.0 to 71.3 and 67.1 for the 5-wall and 7-wall variants, respectively. Matched-wall distance also worsens in the 8-view setting. These results support our choice of a 3-wall MOIS window: a compact local prediction target is easier to learn under sparse wide-baseline observations, while longer ordered wall sequences accumulate more ambiguity from occlusion, room-boundary uncertainty, and distant structural details. The degradation is gradual rather than catastrophic: increasing the MOIS window makes the task harder, but does not break the formulation. This suggests that locality is helpful, with 3 walls giving the best operating point in our setting, while the broader advantage comes from replacing direct global prediction with an anchor-conditioned local representation. Table E.1: MvFFN layout accuracy reported on 5- and 8-view test set from ZInD-XV. The model is trained to predict different numbers of adjacent walls in MOIS prediction. # walls 5-view wall 5-view junction 8-view wall 8-view junction Prec↑ Rec↑ F1↑F1 Dist↓ Prec↑ Rec↑ F1↑F1 Dist↓ Prec↑ Rec↑ F1↑F1 Dist↓ Prec↑ Rec↑ F1↑F1 Dist↓ 3-wall 88.5 87.7 88.1 13.9 79.2 68.2 73.3 13.8 89.4 87.5 88.4 19.0 81.0 69.8 75.0 19.1 5-wall 88.7 82.8 85.6 14.7 78.8 64.2 70.8 17.0 88.4 83.7 86.0 19.6 79.4 64.7 71.3 23.2 7-wall 87.6 81.4 84.4 16.2 76.3 58.7 66.4 18.7 87.7 78.9 83.1 19.2 77.9 59.0 67.1 21.7 E.3 Architecture Ablation Details We compare four architecture variants. The shared-encoder model removes the dedicated semantic feature encoder and predicts MOIS directly from geometry-branch features. The no-fusion model keeps separate geometry and semantic branches but removes geometry-to-semantic feature fusion, while retaining the same downstream Layout Chaining procedure. The one-stage model uses the full two-branch architecture but trains all objectives jointly from scratch. The full model uses separate geometry and semantic branches with the proposed two-stage schedule. The no-fusion variant substantially degrades layout recall, F1F1, and matched-wall distance, indicating that semantic image features alone are insufficient and that geometry-conditioned semantic features are important for cross-view wall grouping and within-room connectivity. The shared-encoder and one-stage variants both become unstable during training: the shared-encoder model collapses after roughly 25k steps, while the one-stage model collapses after roughly 35k steps. This suggests conflicting optimization pressure between dense geometry prediction, which relies on precise one-to-one cross-view matching, and MOIS prediction, which groups columns by wall identity and room structure. Separating the geometry and semantic branches and training them in stages improves both stability and final layout accuracy. Appendix F Dataset Preparation F.1 ZInD-XV Dataset Preparation ZInD-XV is derived from all scenes in the Zillow Indoor Dataset (ZInD) Cruz et al. [2021]. It contains 66,834 panoramas, 167,983 unique wall instances, 23,651 rooms, and 2,706 floors from 1,524 residential house scenes in North America. ZInD-XV retains the original panoramas and room layouts in the global frame, but augments them with globally indexed multi-room floor-plan annotations. Specifically, each floor is represented as a set of room layouts with doors and windows, encoded as closed room polygons over globally unique wall instances in the same format as FloorPlan60K used in Li et al. [2025]. Because ZInD often partitions large open spaces into smaller partial rooms, we recursively merge adjacent rooms connected by openings that do not resemble doorways. As a result, the room partition in ZInD-XV is less fragmented than the original ZInD partition in open-plan areas, while still preserving meaningful room-level structure rather than collapsing each floor into a single geometric region. These global wall instances are further projected into each panorama to produce column-wise floor boundaries and wall-instance semantic labels. Because a single panorama can observe walls from multiple rooms, and visible room connectivity depends on doorway state and view position, these per-panorama labels cannot be obtained by projecting a single room polygon independently. Instead, for each panorama, we search the closest room graph connected through open doors and select the room configuration whose rendered floor boundary best matches the 1D floor boundary predicted by CovisPose Hutchcroft et al. [2022]. This graph is then used to generate the per-panorama floor-boundary and wall-instance labels. Finally, we apply a manual cleanup pass to mask panorama columns with inconsistent floor boundaries or incorrect wall-instance assignments. We plan to release the ZInD-XV annotations, including global floor-plans with unique wall instances, per-panorama floor boundaries, wall-instance semantic labels, and validity masks. F.2 HM3D-XV Dataset Preparation HM3D-XV is constructed from residential scenes in the Habitat--Matterport 3D Semantic Dataset (HM3DSem) Yadav et al. [2023], restricted to scenes with sufficient mesh coverage to better match the residential scene distribution of ZInD-XV. It contains 17,943 panoramas from 372 eligible floors across 180 scenes, split into 287 training floors and 85 validation floors following HM3DSem. Although HM3DSem provides textured house meshes with semantic labels, it does not provide the panorama-based annotations required by our model: camera-posed panoramas, globally indexed wall instances, room-level wall sequences, and per-panorama column-wise floor-boundary and wall-instance labels. We construct these annotations in three stages: automated floor preparation, room-level wall sequence annotation, and per-panorama column-wise annotation. We plan to release the HM3D-XV annotations, including global floor-plans with unique wall instances, per-panorama camera poses, floor boundaries, wall-instance semantic labels, and validity masks. Automated floor preparation. This stage converts each multi-floor HM3DSem scene mesh into per-floor sub-scenes with panoramas of known camera poses and initialized global wall instances. We first segment each scene mesh by floor using the semantic labels to identify floor and ceiling geometry. The floor and ceiling heights used for this partition are estimated from the median vertical positions of these surfaces, and nearby split floor regions are merged with simple geometric heuristics. For each floor, we sample panorama centers from the Habitat navigation mesh with a density-controlled greedy Poisson-disk strategy Bridson [2007], targeting approximately 2 panoramas per square meter at a fixed height of 1.2 m. Using the Habitat-Sim renderer Ramakrishnan et al. [2021], we render 2048×10242048× 1024 RGB equirectangular panoramas from the textured mesh with upright camera orientation. Candidate samples that fall on clutter are discarded, and panoramas with severe visible mesh defects are removed during manual inspection. We then initialize global wall instances at the floor level. Vertical wall meshes are grouped by semantic wall instance label from HM3DSem and projected to the floor plane, and a 2D line is fit to each projected wall instance with RANSAC to recover the wall parameters defined in Li et al. [2025]. Room-level wall sequence annotation. Starting from the initialized global wall instances, this step produces the room-level wall sequences used as the final multi-room floor-plan layout for each floor. We designed a custom graphical interface that lets annotators reorder, merge, and split wall instances, and group them into room-wise ordered sequences with globally consistent wall identities. Each room sequence may be closed or open, since incomplete mesh coverage in HM3DSem often prevents some rooms from being represented as closed polygons. Per-panorama column-wise annotation. Given the annotated multi-room wall layout and panorama camera poses, we project floor geometry and wall instances into each panorama to generate column-wise floor boundaries and wall-instance semantic labels. A separate manual annotation pass is then used to mask out columns with subpar projected assets, for example when floor boundaries are misaligned with visible structure or wall-instance assignments are incorrect. Appendix G Baseline Details We evaluate modular baselines under a common three-stage pipeline: (1) pose estimation in a shared floor-plane frame, (2) construction of pose-aligned floor boundaries or point-maps, and (3) layout vectorization. This setup lets us compare existing pose, boundary, and layout components under the same sparse multi-view input setting. Unlike MvFFN, these baselines do not perform joint multi-view reasoning directly from panorama images in the final layout stage; the layout backend receives only pose-aligned geometric inputs, so upstream pose or boundary errors are already fixed before layout prediction. G.1 Pose Evaluation Pose estimation baselines. We compare against two pose baselines: CovisPose+Greedy Spanning Tree (GST) and VGGT. For CovisPose, pairwise relative panorama poses are predicted for all panorama pairs, and global 3-DoF poses (2D position and 1D rotation) are recovered by chaining relative transforms along a co-visibility-ranked greedy spanning tree, following prior panoramic pose-estimation work Nejatishahidin et al. [2023], Li et al. [2025]. VGGT Wang et al. [2025] does not operate directly on 360∘360 panoramas, so we adapt it through perspective decomposition. We exclude GraphCovis Nejatishahidin et al. [2023] and PanoPose Tu et al. [2024] because public code is unavailable. For each view-count setting, we evaluate on 1,000 N-view test tuples from ZInD-XV. Tuples are generated following the prior pose-evaluation protocol Nejatishahidin et al. [2023]: views are sampled from the same floor with minimum pairwise co-visibility of 0.3, and floors are sampled to maintain broad floor coverage. Predicted pose graphs are aligned to ground truth before evaluation following Nejatishahidin et al. [2023], Li et al. [2025]. We report per-panorama position error in centimeters and yaw error in degrees, summarized by mean, standard deviation, and median over all evaluated panoramas. VGGT panorama adaptation. Since VGGT operates on perspective images, each 360∘360 panorama is decomposed into overlapping rectilinear crops with 120∘120 field of view, 60∘60 horizontal overlap, resolution 560×560560× 560, and pitch angles 0∘0 and −30∘-30 . All crops from all panoramas are processed jointly in a single forward pass. We recover metric scale by aligning the vertical histogram peak of the predicted point cloud to the camera height provided by ZInD Cruz et al. [2021]. We then estimate one planar pose per panorama by compensating each crop pose for its known azimuthal offset and robustly fitting a consensus position and yaw across crops. G.2 Layout Baseline Pipelines Layout baseline pipelines. Given Stage-1 poses, all modular layout baselines reconstruct multi-view layouts by first building geometric inputs in the global floor frame and then applying a layout vectorization backend. Unless otherwise noted, we use CovisPose poses; GT-pose variants replace them with ground-truth poses to isolate layout error from pose error. LGT-Net provides panorama-level floor-wall boundary cues, while RoomFormer and PolyRoom use fused top-down geometric inputs. We do not use VGGT poses for layout baselines because its pose accuracy is substantially worse than CovisPose on ZInD-XV. LGT-Net backend. LGT-Net Jiang et al. [2022] is applied independently to each panorama using the pretrained ZInD checkpoint without fine-tuning. Its predicted floor-wall boundary is converted into a local floor-plan polygon using Manhattan-world post-processing with Douglas–Peucker simplification (ϵ=0.1)(ε=0.1), then projected into the global floor plane using the input poses. Polygons whose overlap exceeds 50%50\% of the smaller polygon area are grouped and merged by Shapely union. We also report a GT-pose variant that replaces predicted poses with ground-truth poses. Topdown training data construction. RoomFormer Yue et al. [2023] and PolyRoom Liu et al. [2024] operate on topdown geometric inputs rather than panoramas, so we finetune both models on ZInD-XV training data constructed to mimic sparse multi-view observations. Starting from lifted floor-plans with fixed ceiling height and room connectivity preserved through open doors, we project per-panorama floor wall boundary points into the global frame using groundtruth poses and rasterize them into 256×256256× 256 orthographic floor plane density maps. The corresponding room polygons from ZInD-XV are used as supervision. To match test-time conditions, we train with 2, 3, 5, and 8-view samples and add Gaussian noise with σ=0.001σ=0.001 on boundary locations, σ=0.1σ=0.1 m on panorama positions, and σ=0.05σ=0.05 rad on panorama yaw. RoomFormer. RoomFormer Yue et al. [2023] is initialized from its Structured3D checkpoint with a ResNet-50 backbone, 800 vertex queries, and up to 20 room polygons per scene. We fine-tune RoomFormer on the rendered ZInD-XV density maps and room polygon supervision described above, using AdamW with learning rate 2×10−42× 10^-4 (2×10−52× 10^-5 for the backbone), weight decay 10−410^-4, batch size 10, and gradient clipping at max norm 0.1. Fine-tuning runs for approximately 70 epochs from the Structured3D initialization within a MultiStepLR schedule whose learning-rate decay is configured at epoch 400 (not reached). At inference, we render the topdown density map from the evaluated input poses and predicted LGT-Net floor wall boundaries, then apply the same vectorization pipeline: corners with foreground probability >0.5>0.5 are retained, polygons with area <100px2<100\,px^2 are discarded, and surviving corners are mapped back to metric floor plane coordinates using a precomputed homography. PolyRoom. PolyRoom Liu et al. [2024] uses the same rendered input representation, but adds a Mask2Former Cheng et al. [2021] room-instance segmentation stage whose predicted room contours initialize a RoomFormer style polygon decoder instead of fixed learned queries. We train PolyRoom in two stages: first fine-tuning a Mask2Former with Swin-L backbone initialized from ImageNet-22K pretraining, then fine-tuning the PolyRoom decoder from its Structured3D checkpoint on the same rendered ZInD-XV data. Stage 1 trains Mask2Former at 256×256256× 256 input resolution with AdamW (learning rate 10−410^-4, weight decay 0.050.05, batch size 2) for 400 epochs, with a backbone learning-rate multiplier of 0.1 and standard random-flip, random-resize, and random-crop augmentations. Stage 2 resumes from the Structured3D-pretrained PolyRoom checkpoint, including its optimizer and scheduler state, and continues fine-tuning on ZInD-XV for the remainder of the original MultiStepLR schedule (configured for 500 total epochs with a learning-rate decay at epoch 400) using AdamW (learning rate 2×10−42× 10^-4, 2×10−52× 10^-5 for the backbone), weight decay 10−410^-4, batch size 10, gradient clipping at max norm 0.1, 800 vertex queries, and up to 20 room polygons per scene. At inference, Mask2Former instances with score >0.6>0.6 are converted to contours using Douglas–Peucker simplification (ϵ=5px)(ε=5\,px) and used to seed the decoder. From the decoded polygons, we retain corners with foreground probability >0.1>0.1 and interior angles within 30∘30 of a right angle, merge nearby corners within 10 pixels, discard polygons with area <100px2<100\,px^2, and map the surviving polygons back to metric floor-plane coordinates. G.3 Layout Evaluation Metrics We evaluate wall and junction precision, recall, and F1F1 using instance-level matching. A predicted wall is considered a candidate match if its orientation differs by at most 10∘10 and its perpendicular distance is within 2.02.0 m of a ground-truth wall. Candidate wall pairs are then matched by Hungarian assignment based on line overlap. Junction precision, recall, and F1F1 are defined analogously: a junction is counted as matched when its two incident walls are both matched. We also report mean matched-wall localization error. We use these structure-aware metrics because MvFFN predicts both closed polygons and open wall sequences, making region-based IoUIoU not well defined for all outputs. G.4 Downstream Refinement Protocol We evaluate planar bundle adjustment (BA) Zhou [2023] and BADGR Li et al. [2025] as downstream refinement methods applied to the MvFFN layouts from Sec. 5.3 of the main paper. For each method, we report mean and median matched wall distances before and after refinement on the full evaluation set, and on the subset of samples whose mean matched wall distance improves; the latter also defines the improvement rate. Before refinement, partial room shapes are closed with a simple rule-based procedure to avoid self-intersections. This can introduce a small number of completed walls without image correspondences, which is especially challenging for BADGR because it assumes the number of room walls is known. This explains why BADGR is more sensitive to initialization in our setting. Appendix H Failure Cases Figure H.1: Failure case caused by incorrect MOIS wall-instance prediction, where MvFFN fails to distinguish different wall instances across views. A first failure mode of MvFFN occurs when the predicted MOIS field does not consistently separate wall instances across views, producing incorrect wall grouping. As shown in Fig. H.1, these errors propagate through layout chaining and lead to both connectivity and geometric mistakes in the final layout. The failure likely stems from limited training coverage for non-Manhattan structures such as bay windows, where cross-view wall-instance prediction is more ambiguous. Although layout chaining suppresses weak wall hypotheses supported by only a few columns, noisy predictions from distant regions can still introduce incorrect connectivity. Inconsistent annotations for structures such as kitchen islands further worsen this problem under limited training data. More generally, because layout chaining is heuristic-based, it has limited ability to recover from noisy MOIS predictions, highlighting the value of a learned post-processing module for robust global layout assembly. Figure H.2: Failure case caused by geometric errors in the initial MvFFN predictions, where inaccurate poses and point maps lead to distorted layout geometry. A second failure mode arises from geometric errors in the initial MvFFN predictions, including inaccurate camera poses and noisy point maps, as shown in Fig. H.2. Because layout chaining builds the global layout from these predicted geometric cues, such errors can propagate to the final reconstruction and produce misaligned walls or distorted room geometry. Although the wall-instance predictions of MvFFN enable downstream bundle adjustment, the quality of the final refinement still depends on the accuracy of the initial geometry and correspondences. Appendix I Limitations and Broader Impact Limitations. MvFFN assumes straightened equirectangular RGB panoramas from a single connected indoor pose graph, with sufficient overlap between neighboring views and known relative camera heights. Performance may degrade when these assumptions are violated, for example under extremely sparse captures, low co-visibility, severe occlusion, inaccurate camera-height metadata, strongly non-Manhattan or curved architecture, low-light or textureless images, or scenes outside the residential indoor domains evaluated here. Because structured multi-room training data remains limited, the current model uses two-view rotation and correspondence priors during training and inference, making performance sensitive to prior quality under domain shift. The method can return partial or open room layouts when visibility is insufficient, and downstream refinement can degrade when the initial layout or pixel-to-wall correspondences are noisy. The dense anchor-conditioned MOIS inference and attention layers also scale with the number of views and image columns, so memory and runtime should be considered when applying the method to larger view sets. Broader impact and responsible use. This work can support useful applications such as lower-cost indoor digital twins, AR/VR walkthroughs, assistive indoor navigation, and robotics mapping. At the same time, accurate reconstruction of private interior layouts can create privacy and security risks if used without consent, and erroneous reconstructions may be harmful in downstream navigation, robotics, or planning systems if treated as safety-critical. We therefore intend the method and released assets for research and consented indoor reconstruction, not for covert mapping, access-control, insurance, credit, law-enforcement, or other consequential decisions. Released annotations and code should follow the licenses and access restrictions of the underlying datasets, document intended use and limitations, and avoid distributing raw indoor imagery beyond what the original dataset terms permit. Appendix J More Qualitative Results J.1 ZInD-XV Results Figure J.1: Qualitative comparisons on ZInD-XV test set. Figs. J.1 and J.2 show additional qualitative examples on ZInD-XV from 5-view and 8-view samples. The first columns visualize the predicted multi-view wall instances with consistent colors across panoramas, followed by the predicted top-down layout, the ground-truth layout, and the strongest modular baseline, LGT-Net with CovisPose. Across both simple rooms and larger multi-room scenes, MvFFN more reliably reconstructs complete room shapes, remains accurate under partial occlusion, and recovers distant wall instances and neighboring room structure visible through doorways. Consistent with the quantitative pose and layout results, the predicted camera poses and wall instances show fewer duplications and omissions than the baseline. In contrast, the baseline often predicts partial room shapes, splits large rooms into smaller fragments, and fails to reason jointly across views, leading to incorrect boundaries, missing wall structure, and redundant contours. Figure J.2: More Qualitative comparisons on ZInD-XV test set. J.2 HM3D-XV Results Figure J.3: Qualitative comparisons on HM3D-XV test set. Figure J.3 shows HM3D-XV reconstruction examples under the furnished, mesh-rendered domain shift. As in ZInD-XV, MvFFN can recover globally consistent wall instances and coherent room-level structure, reducing fragmented or duplicate walls relative to modular geometry-first baselines. Because HM3D-XV is smaller, noisier, and more sparsely annotated, these qualitative results illustrate representative successful cases rather than the full distribution of outcomes. Appendix K Assets, Licenses, and Responsible Release Assets, licenses, and data release. This work uses existing datasets and research artifacts, including ZInD Cruz et al. [2021], HM3D/HM3DSem Yadav et al. [2023], and the baseline or dependency code used in our comparisons. Table K.1 lists the version or access date, citation, role, license or access terms, and release treatment for each asset, including pretrained weights and baseline implementations where applicable. ZInD-XV and HM3D-XV are derived annotation assets. Unless further restricted by the original dataset terms, we release our derived annotations, splits, metadata, and evaluation files under C BY-NC 4.0. This license applies only to the annotations and metadata created in this work, not to the underlying raw panoramas, meshes, textures, renderings, or other source assets. We do not redistribute raw restricted source data; users must obtain ZInD and HM3D/HM3DSem from the original providers and comply with their respective licenses, terms of use, and access restrictions. We also provide conversion and evaluation scripts, together with a dataset card or README documenting label definitions, annotation and conversion pipelines, statistics, splits, file formats, known limitations, intended use, out-of-scope use, and maintenance policy. Asset Role in this work Version / access date License or terms Release treatment ZInD-XV annotations New derived annotation asset v1.0 C BY-NC 4.0, subject to original ZInD terms We release derived wall-instance labels, splits, metadata, conversion scripts, and evaluation files only. HM3D-XV annotations New derived annotation asset v1.0 C BY-NC 4.0, subject to original HM3D and HM3DSem terms We release derived annotations, splits, metadata, conversion scripts, and evaluation files only. Table K.1: Existing and newly derived assets used in this work. We release only assets created by this work under the stated terms and do not redistribute raw restricted source data. Dataset card summary. ZInD-XV contains derived multi-view floor-plan annotations, globally indexed wall instances, room-wise wall sequences, and per-panorama column-wise floor-boundary and wall-instance labels. HM3D-XV contains analogous derived annotations for residential HM3DSem scenes. The released files include train/validation/test splits, per-floor metadata, per-panorama labels, wall-instance identifiers, room connectivity, and evaluation tuples. The README specifies the file format, coordinate convention, label definitions, preprocessing steps, conversion scripts, and evaluation protocol. The intended use is research on indoor layout reconstruction, camera pose estimation, and structured multi-view scene understanding. Out-of-scope uses include covert mapping of private spaces, security-sensitive reconstruction without consent, and consequential decision-making based on predicted layouts. Responsible release. Indoor floor-plans can reveal sensitive information about private spaces. To reduce privacy, security, and license risks, we do not redistribute raw restricted panoramas, meshes, textures, renderings, or floor-plan source files. The released package contains only compatible derived annotations, metadata, splits, and scripts. Users are required to obtain the original datasets from their providers and comply with the corresponding terms. The release documentation states intended uses, out-of-scope uses, limitations, and access restrictions. We will release all code, trained checkpoints, and utilities upon acceptance. Annotation provenance. Manual annotation and cleanup for ZInD-XV and HM3D-XV were performed by the authors/research team using custom annotation and quality-control tools. No crowdworkers, contractors, or human-subject study participants were recruited. The released documentation describes the annotation interface, label definitions, quality-control procedure, and cleanup criteria. Appendix L Reproducibility and Compute Compute resources. All MvFFN experiments were run on 8 or 16 A100 GPUs with 40GB memory per GPU, using 6 12 cpus per gpu. Note that the experiments can also be conducted on lower-tier gpus. Stage-1 geometry pre-training used approximately 2 days for 80k iterations, and stage-2 layout semantic training used approximately 2.5 days for 80k iterations. We used a batch size of [10,7,4,2] for 2, 3, 5, 8 views with 6 12 dataloader workers per process.