Paper deep dive
Where Appearance Fails, Geometry Recognizes: A CAD-Free 3D Shape Prior That Complements Vision Foundation Models
Chenxi Tao, Seung-Kyum Choi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 9/7/2026, 4:58:02 AM
Summary
This paper introduces a CAD-free 3D shape prior for object recognition that complements vision foundation models. By reconstructing objects via 3D Gaussian Splatting (3DGS) from short scans, the authors create geometry prototypes that fuse with frozen DINOv2 image features. The approach recovers the recognition value of CAD models without requiring them, performing comparably to CAD on datasets like HOPE and T-LESS. The geometry prior is particularly effective for shape-distinctive objects and rescues recognition failures under partial occlusion, proving that 3D geometry, not rendered pixels, drives the improvement.
Entities (8)
Relation Signals (6)
Chenxi Tao → affiliatedwith → Georgia Institute of Technology
confidence 99% · Chenxi Tao a , Seung-Kyum Choi a,∗ a George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
Seung-Kyum Choi → affiliatedwith → Georgia Institute of Technology
confidence 99% · Chenxi Tao a , Seung-Kyum Choi a,∗ a George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology
3D Shape Prior → evaluatedon → T-LESS
confidence 98% · geometry from RGB-D depth (on T-LESS), 3DGS, and CAD gives comparable recognition
3D Shape Prior → evaluatedon → HOPE
confidence 98% · on shape-distinctive household objects (HOPE) geometry alone reaches 0.920
3D Shape Prior → fusedwith → DinoV2
confidence 95% · fused with frozen DINOv2 image features
3D Gaussian Splatting → usedfor → 3D Shape Prior
confidence 95% · each object is reconstructed with 3D Gaussian Splatting (3DGS), summarized into a per-class shape prototype
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recognizing specific objects onboarded without a labeled training set recurs across manufacturing and service robotics, yet the conventional renderable prior, a computer-aided-design (CAD) model, is often unavailable. Two-dimensional capture supplies no shape prior, and frozen foundation features fail on geometrically similar, low-texture industrial parts. We ask what a short object-centric scan buys for recognition beyond the captured images themselves: each object is reconstructed with 3D Gaussian Splatting (3DGS), summarized into a per-class shape prototype, and fused with frozen DINOv2 image features. First, the scan recovers the recognition value of CAD without CAD: geometry from RGB-D depth (on T-LESS), 3DGS, and CAD gives comparable recognition (tied on HOPE, within 1.6 points on T-LESS); 3DGS is only a convenient route to a point cloud. Second, the payoff is governed by how recognizable the shape is: on shape-distinctive household objects (HOPE) geometry alone reaches 0.920 versus image-only 0.832, a ceiling below which fixed-weight fusion (0.872) sits. On shape-confusable textureless industrial parts (T-LESS) the gain is modest but consistent (0.560 to 0.591 fused, above both single signals). Third, the prior is complementary, not uniformly additive: it rescues far more image failures than it breaks successes, and its benefit grows under partial occlusion. Finally, the worth lies in geometry, not rendered pixels: 3DGS renderings do not help the image side, and frozen-feature recognition is nearly lighting-invariant (within 2.5 points). The study is scoped to recognition, not the BOP pose benchmark.
Tags
Links
- Source: https://arxiv.org/abs/2609.04381v1
- Canonical: https://arxiv.org/abs/2609.04381v1
Trouble viewing inline? Open PDF directly →
Full Text
64,479 characters extracted from source content.
Expand or collapse full text
Highlights Where Appearance Fails, Geometry Recognizes: A CAD-Free 3D Shape Prior That Comple- ments Vision Foundation Models Chenxi Tao, Seung-Kyum Choi • A short object scan recovers CAD’s recognition value without CAD. • The prior is reconstruction-agnostic: depth (T-LESS), 3DGS, and CAD geometry agree. • The payoff tracks shape distinctiveness: decisive on HOPE, modest on T-LESS. • Complementary prior: rescues image failures, most under partial occlusion. • The recognition value lies in 3D geometry, not rendered pixels or relighting. arXiv:2609.04381v1 [cs.CV] 3 Sep 2026 Where Appearance Fails, Geometry Recognizes: A CAD-Free 3D Shape Prior That Complements Vision Foundation Models Chenxi Tao a , Seung-Kyum Choi a,∗ a George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, 30332, GA, USA A R T I C L E I N F O Keywords: Object recognition CAD-free recognition 3D Gaussian Splatting Vision foundation models 3D shape prior Industrial machine vision A B S T R A C T Recognizing specific objects onboarded without a labeled training set recurs across manufac- turing and service robotics, yet the conventional renderable prior, a computer-aided-design (CAD) model, is often unavailable. Two-dimensional capture supplies no shape prior, and frozen foundation features fail on geometrically similar, low-texture industrial parts. We ask what a short object-centric scan buys for recognition beyond the captured images themselves: each object is reconstructed with 3D Gaussian Splatting (3DGS), summarized into a per-class shape prototype, and fused with frozen DINOv2 image features. First, the scan recovers the recognition value of CAD without CAD: geometry from RGB-D depth (on T-LESS), 3DGS, and CAD gives comparable recognition (tied on HOPE, within 1.6 points on T-LESS); 3DGS is only a convenient route to a point cloud. Second, the payoff is governed by how recognizable the shape is: on shape-distinctive household objects (HOPE) geometry alone reaches 0.920 versus image-only 0.832, a ceiling below which fixed-weight fusion (0.872) sits. On shape-confusable textureless industrial parts (T-LESS) the gain is modest but consistent (0.560 to 0.591 fused, above both single signals). Third, the prior is complementary, not uniformly additive: it rescues far more image failures than it breaks successes, and its benefit grows under partial occlusion. Finally, the worth lies in geometry, not rendered pixels: 3DGS renderings do not help the image side, and frozen-feature recognition is nearly lighting-invariant (within 2.5 points). The study is scoped to recognition, not the BOP pose benchmark. 1. Introduction Industrial and service robots increasingly must recognize specific objects that were never part of a labeled training set, onboarded instead from a quick object-centric capture. CAD models, the conventional renderable prior, are frequently unavailable in practice: supplier-restricted, outdated relative to the physical part, or never authored at all. That absence splits industrial visual sim-to-real into CAD-guided and CAD-unavailable regimes (Tao and Choi, 2026). Without CAD, practitioners fall back on two-dimensional cut-and-paste of captured images (Dwibedi, Misra and Hebert, 2017), which can vary background and pose but cannot supply a three-dimensional shape prior. Frozen foundation features (Oquab, Darcet, Moutakanni et al., 2024) are a powerful alternative, but they can fail on the geometrically similar, low-texture parts common in manufacturing, where appearance alone is weakly discriminative. Recent work fuses geometric and vision-foundation features (Caraffa, Boscaini, Hamza and Poiesi, 2024) or reconstructs objects with Gaussian Splatting for model-free detection (Liu, Wang, Li, Li, Zhang, Huang and Ji, 2024). Template-matching methods built on foundation features have advanced unseen-object detection and pose (Nguyen, Groueix, Ponimatkin, Lepetit and Hodan, 2023; Lin, Liu, Lu and Jia, 2024; Örnek, Labbé, Tekin, Ma, Keskin, Forster and Hodan, 2024). These methods target six-degree-of-freedom pose under the BOP benchmark (Nguyen, Tyree, Guo, Fourmy, Gouda, Lee et al., 2025). How much a scan-derived three-dimensional prior helps recognition, how it compares to CAD, and when it helps remain uncharacterized. We treat a scan-derived 3DGS reconstruction as a geometry prior, a per-class shape prototype, and fuse it with frozen foundation image features for CAD-unavailable recognition. Prior CAD-free Gaussian-Splatting pipelines reuse the reconstructed per-object model for detection under BOP (Liu et al., 2024), and model-based methods match frozen features against a provided mesh for pose (Örnek et al., 2024; Caraffa et al., 2024). We characterize the upstream recognition question instead: on a textureless-industrial dataset (T-LESS) and a textured-household dataset (HOPE), we measure how much the prior helps and what governs its value. We make four contributions. First, the prior is ∗ Corresponding author ctao40@gatech.edu (C. Tao); schoi@me.gatech.edu (S. Choi) ORCID(s): C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 1 of 18 A CAD-Free 3D Shape Prior for Object Recognition reconstruction-agnostic and recovers the recognition value of CAD without CAD on both datasets: geometry from RGB-D depth (on T-LESS), 3DGS, and CAD gives comparable recognition, within seed-level variability. 3DGS is merely a convenient source of a point cloud. Second, the payoff is governed by shape recognizability. On shape- distinctive objects the geometry alone outperforms strong image features; on shape-confusable parts, whose near- identical geometry bounds any shape signal, fusing the prototype with frozen image features still improves recognition, modestly but consistently. Third, the prior is targeted and complementary: it rescues far more image failures than it harms image successes, its benefit is largest where geometry recognizes a shape that appearance confuses, and it grows under partial occlusion. Its worth lies in the reconstructed geometry, not in rendered pixels or relighting. Finally, the effect is not driven by descriptor sophistication: a 45-dimensional training-free descriptor matches both an in-domain PointNet++ and a frozen Objaverse-pretrained 3D foundation encoder, indicating that geometry quality, not encoder capacity, carries the signal. 2. Related Work CAD-free and model-free object recognition and pose The BOP benchmark has steadily shifted from model- based, seen-object tasks toward unseen and model-free settings (Hodaň, Sundermeyer, Drost et al., 2020; Nguyen et al., 2025). A dominant recipe renders views of the object and matches them against the query: with frozen foundation descriptors in CNOS (Nguyen et al., 2023) and SAM-6D (Lin et al., 2024) for unseen-object detection and segmentation and in FoundPose (Örnek et al., 2024) for pose, and with learned template matching or render-and-compare in GigaPose (Nguyen, Groueix, Salzmann and Lepetit, 2024), MegaPose (Labbé, Manuelli, Mousavian, Tyree, Birchfield, Tremblay, Carpentier, Aubry, Fox and Sivic, 2022), OSOP (Shugurov, Li, Busam and Ilic, 2022), OVE6D (Cai, Heikkilä and Rahtu, 2022), and FoundationPose (Wen, Yang, Kautz and Birchfield, 2024) for pose and tracking. This unseen-object line departs from earlier seen-object, per-object-trained regressors such as GDR-Net (Wang, Manhardt, Tombari and Ji, 2021a). OnePose (Sun, Wang, Zhang, He, Zhao, Zhang and Zhou, 2022) onboards objects without CAD from a video sweep, and GFreeDet (Liu et al., 2024) reconstructs objects with Gaussian Splatting for model-free detection. These works target 6D pose or detection under BOP, whose model-free track onboards each object from captured video under a fixed budget, not from a CAD model. Our scan-based onboarding instantiates the same regime, the CAD-unavailable end of the prior-availability spectrum (Tao and Choi, 2026). We reuse their building blocks but ask what the reconstructed geometry contributes to recognition, and how it compares to CAD. Geometry and vision-foundation fusion Closest to our mechanism, FreeZe (Caraffa et al., 2024) fuses a frozen point-cloud encoder with frozen vision features for training-free pose. Cross-modal methods learn a joint image–point- cloud embedding, including ULIP (Xue, Gao, Xing, Martín-Martín, Wu, Xiong, Xu, Niebles and Savarese, 2023) and OpenShape (Liu, Shi, Kuang, Zhu, Li, Han, Cai, Porikli and Su, 2023). We share the premise that geometry complements vision-foundation features, but we target classification and obtain the geometry from a scan, with no CAD model. Frozen foundation features are extracted per image and are not inherently view-consistent, which is why template-matching pipelines render and match many views (Nguyen et al., 2023; Lin et al., 2024; Örnek et al., 2024). The per-class geometry prototype supplies the view-consistent 3D structure that the image features alone lack. Radiance fields and Gaussian Splatting as object representations Neural radiance fields (Mildenhall, Srini- vasan, Tancik, Barron, Ramamoorthi and Ng, 2020) and their accelerated, unbounded-scene, and surface vari- ants (Müller, Evans, Schied and Keller, 2022; Barron, Mildenhall, Verbin, Srinivasan and Hedman, 2022; Wang, Liu, Liu, Theobalt, Komura and Wang, 2021b) reconstruct objects from posed images. 3D Gaussian Splatting (Kerbl, Kopanas, Leimkühler and Drettakis, 2023) enables real-time, high-fidelity reconstruction, and surface-oriented variants such as 2D Gaussian Splatting (Huang, Yu, Chen, Geiger and Gao, 2024) and SuGaR (Guédon and Lepetit, 2024) extract meshes. 3DGS has also been used directly for 6DoF pose: novel-object pose (GS-Pose (Cai, Heikkilä and Rahtu, 2025)), online pose tracking (6DOPE-GS (Jin, Prasad, Jauhri, Franzius and Chalvatzaki, 2025)), and camera pose from a single image against a 3DGS model (6DGS (Bortolon, Tsesmelis, James, Poiesi and Del Bue, 2024)). Reconstructed per-object twins underpin model-free onboarding: FoundationPose builds a neural twin from reference images (Wen et al., 2024), and GFreeDet drives the BOP model-free detection track with a 3DGS twin (Liu et al., 2024). Our question sits upstream of these pose tasks: we use 3DGS only to obtain a CAD-free object point cloud, from which we derive a shape prototype. Reusing the same twin for pose is left to future work. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 2 of 18 A CAD-Free 3D Shape Prior for Object Recognition Foundation visual features Self-supervised vision transformers (Caron, Touvron, Misra, Jégou, Mairal, Bojanowski and Joulin, 2021; Oquab et al., 2024; Siméoni et al., 2025) learn strong dense features, aided by register tokens (Darcet, Oquab, Mairal and Bojanowski, 2024); alternative objectives include masked autoencoding (He, Chen, Xie, Li, Dollár and Girshick, 2022), joint-embedding prediction (Assran, Duval, Misra, Bojanowski, Vincent, Rabbat, LeCun and Ballas, 2023), and multimodal autoregression (Fini et al., 2025). Image–text contrastive encoders (Radford, Kim, Hallacy et al., 2021; Zhai, Mustafa, Kolesnikov and Beyer, 2023; Tschannen et al., 2025; Xu, Xie, Tan, Huang, Howes, Sharma, Li, Ghosh, Zettlemoyer and Feichtenhofer, 2024) and promptable segmentation (Kirillov, Mintun, Ravi et al., 2023) are also widely used as frozen backbones. We benchmark 14 such encoders as the frozen image side and find self-supervised vision transformers strongest for textureless industrial parts. A complementary line instead induces a shape bias for textureless and metallic parts, for example by randomizing surface texture during training (Hönig, Thalhammer, Weibel, Hirschmanner and Vincze, 2025). 3D shape descriptors and point-cloud encoders Classical global shape signatures such as shape distributions (Os- ada, Funkhouser, Chazelle and Dobkin, 2002) and learned encoders (Qi, Su, Mo and Guibas, 2017a; Qi, Yi, Su and Guibas, 2017b) summarize a point cloud into a vector. Our geometry prototype uses a compact handbuilt descriptor; we also evaluate a learned encoder. Synthetic data for CAD-free training, and prototype recognition Without CAD, training data is commonly synthesized by cut-and-paste of captured crops (Dwibedi et al., 2017), domain randomization (Tobin, Fong, Ray, Schneider, Zaremba and Abbeel, 2017), physically based rendering (Denninger, Winkelbauer, Sundermeyer, Boerdijk, Knauer, Strobl, Humt and Triebel, 2023), and copy-paste augmentation (Ghiasi, Cui, Srinivas, Qian, Lin, Cubuk, Le and Zoph, 2021); all supply appearance but no 3D shape prior. Our scoring builds on prototype-based few-shot recognition (Snell, Swersky and Zemel, 2017). Classical CAD-free shape priors Before learned features, CAD-free shape cues were recovered from multi-view geometry: the visual hull and shape-from-silhouette reconstruct a coarse 3D shape from object silhouettes (Laurentini, 1994). Separately, edge- and template-based matching recognized textureless objects from many rendered or captured views (Hinterstoisser, Lepetit, Ilic, Holzer, Bradski, Konolige and Navab, 2012). Edge-based cues remain a staple for recognizing textureless parts (Clement, Shah, Pancholi, Lugo Bustillo and Cheng, 2024). We revisit this idea of geometry-from-views without CAD. Instead of matching hand-engineered templates, we pair a modern reconstruction (3DGS or depth) and a compact shape descriptor with frozen foundation image features. Benchmarks We evaluate on the textureless-industrial T-LESS (Hodaň, Haluza, Obdržálek, Matas, Lourakis and Zabulis, 2017) and the textured-household HOPE (Tyree, Tremblay, To, Cheng, Mosier, Smith and Birchfield, 2022). Related object datasets include YCB-Video (Xiang, Schmidt, Narayanan and Fox, 2018), LineMOD (Hinterstoisser et al., 2012), the industrial ITODD (Drost, Ulrich, Bergmann, Härtinger and Steger, 2017), and IMD, a recent benchmark for industrial metallic objects (Ma, Zudaire, Li and Zhang, 2025). 3. Method Task and overview We address CAD-free instance recognition: a query crop must be assigned to one of 퐶 object classes that were onboarded only from a short object-centric scan, with no labeled training images and no CAD model. Per class, the pipeline (Fig. 1) builds two prototypes from the onboarding scan: a frozen-foundation image prototype and a scan-derived geometry prototype. A query is then scored by a weighted sum of its similarities to the two. The backbone is never fine-tuned; the only trainable component is a small head that maps image features into the geometry- descriptor space. Reconstruction Each object is reconstructed CAD-free from its onboarding scan with 3DGS (Kerbl et al., 2023). We train the splatfacto-big model in Nerfstudio for 20,000 iterations on T-LESS and 15,000 on HOPE, with a white background and point-cloud initialization, from the known object-centric camera poses given by the BOP scene_gt annotations (object rotation 푅 푚2푐 and translation 푡 푚2푐 , converted to camera-to-world). We export the Gaussian centers as a point cloud and crop it to the object core (points within 0.15 m of the origin), followed by statistical outlier removal (Open3D, 20 neighbors, standard-deviation ratio 2.0). This discards background drift and floaters, a known 3DGS failure mode under limited or low-texture views (Xiong, Muttukuru, Xiao, Upadhyay, Chari, Zhao and Kadambi, C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 3 of 18 A CAD-Free 3D Shape Prior for Object Recognition (A) Input Onboarding scan (CAD-free) camera frusta orbit turntable ~160 frames / object, no CAD (B) Appearance lane Frozen DINOv2-giant input Transformer Transformer frozen weights 1536-D image prototype (class mean, ℓ 2 ) (C) Geometry lane 3D Gaussian Splatting Point cloud Geometric feature extraction Shape descriptor (45-D) Geometry prototype λ 1 ≥ λ 2 ≥ λ 3 principal feature extraction splatfacto, ~20k iters outlier / floater removal 3 bbox & 3 eigenvalue ratios, 3 bbox stats, 3×12-bin hist. (standardized + ℓ 2 ) (D) Image→geometry head only trained module trainable test-time query: one RGB image; geometry score needs no 3D capture (E) Fusion s c = s c img + λ · s c geom λ = 0.35 ; τ img = 0.01 , τ geom = 0.10 per class c query cos( · , ) / τ im g s c img cos( · , ) / τ g eo m s c geom × λ 1536-D image feature 45-D geometry descriptor (F) Recognition top-1 result obj A 95% obj B 5% under partial occlusion, geometry's edge grows look-alike parts: appearance confuses them, geometry tells them apart Figure 1: Method overview. From a short object-centric onboarding scan we build, per class, two prototypes: a frozen DINOv2-giant image prototype (appearance branch) and a CAD-free geometry prototype from a 3DGS reconstruction (geometry branch). A small head maps image features into the geometry-descriptor space; a query is recognized by the fused similarity 푠 푐 = 푠 img 푐 + 휆푠 geom 푐 . 2025), and leaves a clean per-class point cloud. The crop is automatic and identical for every object (a fixed radius threshold plus the same outlier parameters), so no per-object tuning is required. Each reconstruction takes on the order of a few minutes on a single GPU (NVIDIA RTX 5000 Ada). Figure 2 shows representative reconstructions against the real crops. Geometry prototype Let 푃 = 푝 푖 푁 푖=1 ⊂ℝ 3 be the reconstructed point cloud, centered (푝 푖 ←푝 푖 − ̄푝) and scaled by its maximum radius (푝 푖 ← 푝 푖 ∕ max 푗 ‖푝 푗 ‖). From the covariance 퐶 = 1 푁 ∑ 푖 푝 푖 푝 ⊤ 푖 with eigenvalues 휆 1 ≥ 휆 2 ≥ 휆 3 and eigenvectors 푉 = [푣 1 ,푣 2 ,푣 3 ] we form the eigenvalue ratios 휆 푘 ∕ ∑ 푗 휆 푗 , and in the principal-axis frame 푞 푖 = 푉 ⊤ 푝 푖 the radial coordinate 푟 푖 = ‖(푞 푥 푖 ,푞 푦 푖 )‖ and the height 푧 푖 = 푞 푧 푖 . The 45-dimensional descriptor concatenates three bounding- box edge-length ratios, the three eigenvalue ratios, three bounding-box statistics (product, mean, and standard deviation of the extents), and three normalized 12-bin histograms, of 푟 푖 , of 푧 푖 , and of the pairwise distances ‖푝 푖 − 푝 푗 ‖ over 4,096 sampled pairs (9+3×12 = 45). Writing 푑 푐 ∈ℝ 45 for this descriptor, we standardize it per dimension (with the across-class mean 푚 and standard deviation 휎) and 퓁 2 -normalize to form the per-class prototype, 푔 푐 = ̃ 푑 푐 ‖ ̃ 푑 푐 ‖ , ̃ 푑 푐 = (푑 푐 − 푚) ⊘ 휎,(1) where⊘ denotes elementwise division. The descriptor is largely invariant to rigid motion: the eigenvalue ratios and the pairwise-distance histogram are exact rotation- and translation-invariants (they do not depend on 푉 ), and the principal- axis radial and height histograms are rotation-invariant up to the eigenvector-sign ambiguity. Two singularities remain. When eigenvalues are nearly equal (휆 푎 ≈휆 푏 ), the eigenvectors are not unique, so the principal-axis frame is ambiguous within the degenerate eigenspace and the radial and height histograms can swap or rotate; this affects near-symmetric parts. The axis-aligned bounding-box ratios are the only genuinely pose-dependent terms (coarse size cues); computing them in the principal-axis frame instead would make the descriptor exactly SE(3)-invariant. The components that carry most of the discriminative signal, the eigenvalue ratios and the pairwise-distance histogram, are immune to both singularities. We keep the descriptor handbuilt, not learned, because the CAD-free setting provides neither a CAD mesh nor a large labeled point-cloud corpus to train a point encoder; we verify this empirically against an in-domain PointNet++ and a frozen pretrained 3D foundation encoder in Section 4. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 4 of 18 A CAD-Free 3D Shape Prior for Object Recognition Real crop 3DGS render Point cloud Real crop Point cloud T-LESS (textureless industrial) HOPE (textured household) Figure 2: Reconstruction quality on both datasets, six objects each. Top (T-LESS, textureless industrial): a real onboarding crop, a 3DGS rendered novel view, and the extracted point cloud that feeds the shape descriptor. Bottom (HOPE, textured household): a real crop and the extracted point cloud (geometry-only clouds, shown in neutral gray; 3DGS renders are omitted as the rendered pixels are not used downstream). The CAD-free reconstruction recovers object shape across both textureless parts and textured products; the T-LESS renders also show the render-versus-real domain gap that makes 3DGS a useful geometry source yet an unhelpful substitute image source (Section 4, R5). Image prototype Onboarding crops 푥 ∈ 푂 푐 of class 푐 are encoded by a frozen DINOv2-giant backbone 푓(⋅) (Oquab et al., 2024), giving 1536-dimensional features. The per-class image prototype is their 퓁 2 -normalized mean, 휇 img 푐 = ̄ 푓 푐 ‖ ̄ 푓 푐 ‖ , ̄ 푓 푐 = 1 |푂 푐 | ∑ 푥∈푂 푐 푓(푥),(2) and a query is scored against class 푐 by temperature-scaled cosine similarity, 푠 img 푐 = 1 휏 img ⟨푓(푥), 휇 img 푐 ⟩ ‖푓(푥)‖ ‖휇 img 푐 ‖ , 휏 img = 0.01.(3) Fusion head and scoring A lightweight head ℎ(⋅) maps a 1536-dimensional image feature into the 45-dimensional geometry-descriptor space through Linear(1536→512), ReLU, dropout 0.1, and Linear(512→45), followed by 퓁 2 - normalization so ‖ℎ(푓(푥))‖ = 1. The geometry score of class 푐 is the temperature-scaled cosine between the mapped feature and the geometry prototype 푔 푐 , 푠 geom 푐 = 1 휏 geom ⟨ℎ(푓(푥)), 푔 푐 ⟩, 휏 geom = 0.10.(4) C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 5 of 18 A CAD-Free 3D Shape Prior for Object Recognition With the backbone frozen, only the head is trained, on the onboarding features, by a cross-entropy term over the geometry scores plus a 0.25-weighted alignment term that pulls each mapped feature toward its own class prototype 푔 푦 , = CE ( 푠 geom 푐 푐 , 푦 ) + 0.25 ( 1 − ⟨ℎ(푓(푥)), 푔 푦 ⟩ ) ,(5) optimized with AdamW (learning rate 10 −3 , weight decay 10 −4 , batch size 64) for 80 epochs, separately per dataset. The random seeds control the head initialization, the training shuffling, and the descriptor’s pairwise-distance sampling; the frozen image side is deterministic, which is why image-only results carry no seed deviation. Overfitting on the small number of classes is limited by design: the backbone is frozen, the head is small and maps into a low-dimensional (45) target, and dropout and weight decay regularize it. A query is finally scored by the fusion 푠 푐 = 푠 img 푐 + 휆푠 geom 푐 ,(6) where휆 is a single fixed weight shared across both datasets (no per-dataset tuning). Setting휆=0 recovers the image-only baseline and large휆 recovers geometry-only classification; we use휆 = 0.35. The weight and the two temperatures were fixed once during early development and held unchanged for every experiment, dataset, and backbone; the post-hoc sweep in Fig. 10 confirms the reported setting is not a swept optimum. Recognition is zero-shot in that prototypes come only from the onboarding scan, with no per-class fine-tuning on labeled query images. Datasets and splits We evaluate on T-LESS (Hodaň et al., 2017) (28 of its 30 objects; objects 19 and 20 are excluded because their scans did not yield usable reconstructions, an exclusion fixed at reconstruction time, before any recognition evaluation; 160 onboarding frames per class) and HOPE (Tyree et al., 2022) (28 objects; 60 onboarding frames per class). Recognition queries are real, labeled crops drawn from scenes disjoint from the onboarding scan, which prevents onboarding-to-query leakage: 6,395 queries for T-LESS and 915 for HOPE. Concretely, T-LESS onboarding uses the object-centric train_primesense turntable scans (one object per scene) and queries are crops from the multi-object test_primesense scenes; HOPE onboarding uses the BOP onboarding_static scans and queries are crops from the validation scenes. All query crops are taken at the ground-truth annotation boxes, isolating recognition from detection. Accuracies are top-1 throughout, averaged over 5 random seeds unless noted (3 seeds for the RGB-D depth source in R2). 4. Experiments Setup Datasets, onboarding scans, and the scene-disjoint query protocol are those of Section 3 (“Datasets and splits”). Unless stated otherwise the backbone is a frozen DINOv2-giant, the fusion weight is fixed at 휆 = 0.35, and we report top-1 accuracy averaged over 5 seeds. Prototypes are built only from the onboarding scan, and the only trained component, the fusion head, is trained only on onboarding features; no labeled query images are used anywhere. R1. A CAD-free 3DGS geometry prior improves recognition On T-LESS, fusing the 3DGS-derived geometry prototype with frozen image prototypes raised accuracy from 0.560 (image-only) to 0.591±0.004 (fused, 5 seeds). On HOPE the geometry prototype alone reached 0.920 ± 0.006, far exceeding image-only 0.832; geometry is so dominant here that geometry-only is the strongest single configuration, while fixed-휆 fusion (0.872 ± 0.002) sits between image and geometry (Fig. 3, Table 1). The prior helps in both regimes: as a complement on the shape-confusable industrial parts, and as the primary signal on the shape-distinctive household objects. This yields a simple operational rule, made precise in the weight-robustness analysis below: fuse the two signals where appearance is competitive (T-LESS), and let geometry lead where it already dominates (HOPE). Either way, the prior is best read as a geometry signal, used in fusion or, where geometry suffices, on its own. In practice the regime can be identified at onboarding time, without labeled queries: held-out onboarding frames, labeled by construction, give a proxy estimate of each arm’s accuracy, and the class separability (silhouette) of the frozen image features over the onboarding crops signals whether appearance is competitive (cf. the feature-manifold analysis below). R2. The prior is reconstruction-agnostic: depth, 3DGS, and CAD geometry agree On the same held-out T-LESS test set, geometry from three different sources gives closely comparable recognition: a point cloud back- projected from the onboarding RGB-D depth (geometry-only 0.575±0.005, fused 0.593±0.001; 3 seeds), our 3DGS C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 6 of 18 A CAD-Free 3D Shape Prior for Object Recognition T-LESS (textureless) HOPE (textured) 0.0 0.2 0.4 0.6 0.8 1.0 Recognition accuracy (top-1) 0.560 0.832 0.566 0.920 0.591 0.872 Image-onlyGeometry-onlyFused Figure 3: Recognition accuracy with a frozen DINOv2-giant backbone (휆 = 0.35, 5 seeds). The 3DGS-derived geometry prior helps on both datasets; on the shape-distinctive HOPE objects the geometry alone exceeds strong image features. reconstruction (0.566±0.013 / 0.591±0.004), and surface-sampled CAD (0.582±0.018 / 0.588±0.004). The three sources differ by at most 1.6 points in geometry-only accuracy (within one seed-level standard deviation) and by at most 0.5 points when fused. What matters is the object geometry itself, not any 3DGS-specific property: depth or multi-view stereo would be expected to serve equally well, and we use 3DGS only as a convenient CAD-free route to a point cloud. (Depth is available for T-LESS but not for the HOPE static onboarding scans, so the three-way comparison is on T-LESS; on HOPE, 3DGS and CAD geometry are likewise tied, 0.920 vs 0.916, per-seed differences straddling zero.) In this sense a short scan recovers the recognition value of CAD without CAD. Table 1 Main recognition results (top-1 accuracy, frozen DINOv2-giant, 휆 = 0.35; mean±std over 5 seeds, same held-out test set for all columns). Image-only is seed-independent (deterministic prototypes), so no deviation is reported for it. DatasetObject typeImage-only Geom (3DGS)FusedGeom (CAD) T-LESS textureless, shape-confusable0.5600.566±0.013 0.591±0.004 0.582±0.018 HOPEtextured, shape-distinctive0.8320.920±0.006 0.872±0.002 0.916±0.007 R3. The prior is complementary, helping where appearance fails We stratified each query into image-wrong and image-right cases. On HOPE the geometry rescued 26% of image failures while harming only 0.4% of image successes; on T-LESS it rescued 14% and harmed 5.6% (Fig. 4, Table 2). Fig. 5 shows representative T-LESS rescues, where geometry corrects a confident appearance error on a look-alike part. The benefit is not uniform across classes. Controlling for per-class image difficulty, a class gains in proportion to how well the geometry prototype alone recognizes it (Pearson partial correlation between fusion gain and geometry-only accuracy, controlling for per-class image-only accuracy, 푛=28 classes per dataset: +0.82 on T-LESS, 푝 < 0.001, and +0.37 on HOPE, 푝 = 0.06, the latter suggestive rather than significant at this class count; Fig. 6). A purely geometric distinctiveness metric (inter-prototype C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 7 of 18 A CAD-Free 3D Shape Prior for Object Recognition descriptor distance) did not predict the gain within either dataset (Pearson푟 = +0.06 on T-LESS, −0.15 on HOPE). The effect is one of appearance–geometry complementarity (geometry helps where it can recognize a shape that appearance confuses), not raw shape salience. The cleaner fused confusion diagonal on HOPE than on T-LESS (Fig. 7) reflects HOPE’s larger appearance–geometry gap. T-LESSHOPE 0 5 10 15 20 25 30 Per-query rate (%) 14% 26% 5.6% 0.4% Rescue (image-wrong → fused-right) Harm (image-right → fused-wrong) Figure 4: Per-query effect of the geometry prior relative to the image-only baseline. On HOPE, where shapes are distinctive, the prior rescues about a quarter of image failures with almost no harm; on the shape-confusable T-LESS parts it rescues fewer and harms more. Table 2 Per-query stratification of the geometry prior relative to image-only. Rescue = fraction of image-wrong queries fusion corrects; harm = fraction of image-right queries fusion breaks. DatasetRescue rate Harm rate T-LESS (shape-confusable) 14.2% ± 0.7 5.6% ± 0.4 HOPE (shape-distinctive)26.2% ± 0.8 0.4% ± 0.1 A feature-manifold view of when fusion helps The harm-rate asymmetry (5.6% on T-LESS vs 0.4% on HOPE) follows from how separable the frozen image features are. A t-SNE projection and silhouette analysis of the DINOv2 features (Fig. 8) reveal a far more collapsed manifold on textureless T-LESS (silhouette 0.02) than on textured HOPE (silhouette 0.17): on T-LESS many classes alias into overlapping clusters. Where the appearance manifold is collapsed, even correct image predictions are marginal, so fusing geometry both rescues many failures and occasionally overturns a borderline-correct one (higher harm). Where it is well separated (HOPE), correct predictions are confident and geometry almost never harms them. The geometry prior acts orthogonally to appearance, separating classes where the appearance manifold has collapsed. R4. The benefit grows under partial occlusion If the prior helps most where the appearance manifold collapses, occlusion should be such a regime: it removes appearance evidence in a controlled, measurable way. On HOPE, where BOP visibility annotations are available, recognition degrades gracefully as objects become occluded (fused accuracy 0.93 at full visibility down to 0.71 below 50% visible). The geometry prior’s advantage over appearance grows with occlusion: the per-query fused-over-image gap rises from +2.5 points on fully visible objects to +6–8 points under C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 8 of 18 A CAD-Free 3D Shape Prior for Object Recognition Query obj 6 × Image-only pick obj 14 ×obj 3 ×obj 22 ×obj 3 × obj 11 ✓ Correct (fused) obj 17 ✓obj 4 ✓obj 23 ✓obj 2 ✓ Figure 5: Qualitative rescue examples on T-LESS. Each query (top) is confidently misclassified by the image-only prototype as an appearance-similar object (middle, red) but correctly recognized once the geometry score is fused (bottom, green); representative onboarding crops are shown for the predicted and true classes. The geometry prototype disambiguates parts that look alike to frozen appearance features. heavy occlusion (Fig. 9a). Occlusion is the regime where appearance features fail and a shape prior compensates, which is directly relevant to cluttered, partially occluded bin-picking. On T-LESS, where the same BOP visibility annotations are available for the test scenes (Fig. 9b), the gap follows the same trend through the partial-occlusion range: the fused-over-image gain rises from +3.5 points on near-fully- visible objects (≥ 90%) to +5.8 points at 70–90% visibility. Unlike HOPE, however, it shrinks to +0.4 points under heavy occlusion (<50% visible), where all three signals collapse to a common low level (0.26–0.27). This difference is consistent with the datasets’ content: HOPE objects stay shape-distinctive even when largely hidden, whereas T- LESS’s 28 near-identical industrial parts become geometrically ambiguous once most of the object is occluded, so the partial-cloud descriptor can no longer disambiguate them. Geometry thus helps most in the partial-occlusion regime on both datasets, but it is not a remedy for near-total occlusion, where the shape signal itself is destroyed. Robustness to the fusion weight The preceding results all use a single fixed fusion weight; we now check that this choice is not a tuned artifact. Accuracy rises smoothly and monotonically with the fusion weight 휆 across the swept range 휆 ∈ [0, 1] on both datasets, with no instability (Fig. 10); the reported 휆 = 0.35 is a single conservative value shared across both datasets, not a per-dataset tuned optimum. The best swept weight (휆 = 1) bounds what fusion reaches within the sweep: 0.598 on T-LESS (versus 0.591 at the fixed weight) and 0.897 on HOPE (versus 0.872). On T-LESS, fusion at any reasonable weight exceeds both image-only (0.560) and geometry-only (0.566), confirming genuine complementarity. On HOPE, geometry-only (0.920) is instead the ceiling: fusion approaches it as 휆 grows but cannot exceed it, because geometry strictly dominates appearance for these objects. The under-performance of fixed-휆 fusion on HOPE is intrinsic to this geometry-dominant regime; a per-class adaptive weight, geometry-heavy where geometry already recognizes the class (R3), would recover the geometry-only result. We keep a single fixed weight to avoid per-dataset or per-class tuning. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 9 of 18 A CAD-Free 3D Shape Prior for Object Recognition 0.00.20.40.60.81.0 Geometry-only per-class accuracy −0.2 −0.1 0.0 0.1 0.2 Fusion gain (fused − image) dashed: partial trend (geom | image fixed) partial r: T-LESS +0.82, HOPE +0.37 T-LESS HOPE 0.0 0.2 0.4 0.6 0.8 1.0 image-only accuracy Figure 6: Complementarity of the geometry prior. Each point is a class; color encodes per-class image-only accuracy. The fusion gain is largest for classes that geometry recognizes well (high 푥) but appearance does not (darker points); controlling for image difficulty, the partial correlation between gain and geometry-only accuracy is +0.82 (푝 < 0.001, T-LESS) / +0.37 (푝 = 0.06, suggestive, HOPE). A raw geometric-distinctiveness metric did not predict gain, ruling out shape salience alone. Predicted True (a) T-LESS (fused), 28 classes Predicted True (b) HOPE (fused), 28 classes 0.0 0.2 0.4 0.6 0.8 1.0 row-normalized rate Figure 7: Row-normalized confusion matrices of the fused classifier. The diagonal is far cleaner on HOPE (right) than on T-LESS (left), where residual off-diagonal mass concentrates among geometrically similar classes. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 10 of 18 A CAD-Free 3D Shape Prior for Object Recognition (a) T-LESS feature silhouette = 0.023 (b) HOPE feature silhouette = 0.172 Figure 8: t-SNE of frozen DINOv2 features (equal per-class subsamples; perplexity 30, PCA initialization; silhouette computed with cosine distance in the raw 1536-D feature space, not the 2-D embedding). The textureless T-LESS manifold (left, silhouette 0.02) is far more collapsed (classes overlap) than the textured HOPE manifold (right, silhouette 0.17), explaining why appearance fails more on T-LESS and why the geometry prior both rescues and occasionally harms more there. R5. The worth lies in geometry, not rendered pixels Having established when and how much the prior helps, we ask what about the reconstruction carries the signal. Using 3DGS rendered views as the image source did not help. Replacing the real onboarding crops with renders dropped image-only accuracy from 0.560 to 0.392, and adding renders on top of the real crops slightly hurt it (0.560 to 0.554), consistent with the domain gap between clean renders and cluttered real queries that is visible in Fig. 2; Fig. 11a quantifies the drop. We also probed lighting sensitivity in a controlled way, rendering the 28 T-LESS CAD models under matched, shifted, and extreme lighting conditions that vary key-light direction, color temperature, intensity, and ambient level (24 views per object) and evaluating them by few-shot prototype recognition. The probe left frozen-feature recognition essentially unchanged, within 2.5 points even under extreme illumination (Fig. 11b), so the relighting capability of a renderable asset is inert for this task. What the reconstruction contributes, then, is geometry, not pixels a renderer can synthesize. R6. Backbone study A remaining question is whether these findings are an artifact of the particular image encoder. We benchmarked the same 14 frozen image encoders as the image side on both datasets, all on the same held-out test sets over 5 seeds (Tables 3 and 4). On T-LESS, no single encoder dominates: on fused accuracy the leaders are image–text models (SigLIP2-g 0.629, MetaCLIP-H 0.613), with SigLIP-L (0.592) and DINOv2-giant (0.591) close behind and within a few seeds’ standard deviation of each other; on image-only accuracy the strongest encoders cluster tightly near 0.56 (SigLIP2-g, DINOv2-giant, MetaCLIP-H, SigLIP-so400m (Zhai et al., 2023)). The image–text CLIP- large (Radford et al., 2021) and the masked-prediction I-JEPA (Assran et al., 2023) are markedly weaker (fused 0.36 and 0.15). On HOPE, whose objects are textured and appearance-rich, image-only accuracy is far higher and more spread out (0.80–0.96, excluding the I-JEPA outlier at 0.155), and the leaders shift to OpenCLIP-H (0.966), DINOv3-H+ (0.963), and OpenCLIP-bigG (0.956); DINOv2-giant sits mid-pack (0.872). One pattern is robust across both datasets: the geometry-fusion gain holds for nearly every one of the 14 backbones (fused exceeds image-only for 13 of 14 on T-LESS and all 14 on HOPE; the lone exception, SigLIP-so400m on T-LESS, is 0.003 below, within seed noise). The gain also grows as the backbone weakens: on HOPE it is only +0.010 for the strongest encoder (OpenCLIP-H) but +0.040 for DINOv2-giant and +0.118 for I-JEPA, mirroring at the backbone C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 11 of 18 A CAD-Free 3D Shape Prior for Object Recognition <0.5 (n=48) 0.5-0.7 (n=119) 0.7-0.9 (n=263) 0.9-1.0 (n=485) Object visibility (visible fraction) 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Recognition accuracy (a) HOPE (textured) Geometry-only Fused Image-only <0.5 (n=921) 0.5-0.7 (n=635) 0.7-0.9 (n=1080) 0.9-1.0 (n=3759) Object visibility (visible fraction) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 (b) T-LESS (textureless) Figure 9: Recognition accuracy versus object visibility, stratified by BOP visible fraction, on (a) HOPE and (b) T-LESS (note the different 푦-axis ranges; HOPE accuracies are far higher). Arrows mark the fused-over-image gap in each bin. On HOPE the geometry prior helps most under heavy occlusion; on T-LESS the gain peaks at moderate visibility and collapses below 50% visible, where the 28 near-identical parts become geometrically ambiguous and all signals fall to a common low level. 0.00.20.40.60.81.0 Geometry fusion weight λ 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 Recognition accuracy (5 seeds) reported λ = 0.35 HOPE T-LESS Figure 10: Sensitivity to the geometry fusion weight 휆 (5 seeds, shaded ±std). Accuracy increases monotonically and smoothly on both datasets; the dashed line marks the reported 휆 = 0.35. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 12 of 18 A CAD-Free 3D Shape Prior for Object Recognition real (baseline) real+ renders renders only 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Image-only accuracy (T-LESS) 0.560 0.554 0.392 (a) 3DGS renders do not help matchedshiftedextreme 0.800 0.825 0.850 0.875 0.900 0.925 0.950 0.975 1.000 Recognition accuracy 0.90 0.95 0.90 0.94 0.88 0.92 (b) Lighting-invariant ( ≤ 2.5 pt) 5-shot10-shot Figure 11: The prior’s value is geometric, not photometric. (a) Using 3DGS renders as the image source hurts recognition relative to real crops. (b) A lighting-sensitivity probe: frozen DINOv2 recognition is essentially unchanged from matched to extreme illumination (≤ 2.5 points). level the headline finding that geometry helps most where appearance is least reliable. The benefit is not specific to any one encoder. We adopt DINOv2-giant for all other experiments. It is the de-facto frozen backbone of the unseen- object recognition line we build on (e.g., CNOS (Nguyen et al., 2023), SAM-6D (Lin et al., 2024), FoundPose (Örnek et al., 2024)), which keeps our results directly comparable. It is also fully self-supervised, with no text supervision, in keeping with the CAD-free setting, and its image-only accuracy ties the best on T-LESS. It is, however, not the strongest backbone on HOPE, where the leading image–text encoders would raise fused accuracy from 0.872 to about 0.96, so the HOPE numbers we report elsewhere are conservative. Changing the backbone shifts the absolute level but not the fusion gain or any conclusion. R7. Handbuilt vs learned geometry encoders R6 varied the image side; we now vary the geometry side and ask whether the training-free descriptor is a limitation. The two learned alternatives are an in-domain PointNet++ trained on the same reconstructed clouds (with heavy SO(3)-rotation, jitter, scale, and partial-view augmentation to compensate for having only 28 shapes), and a frozen pretrained 3D foundation encoder, OpenShape (Liu et al., 2023) (a PointBERT (Yu, Tang, Rao, Huang, Zhou and Lu, 2022) point encoder aligned to an OpenCLIP ViT- bigG-14 embedding space, trained largely on Objaverse), used zero-shot exactly as we use frozen DINOv2 on the image side. Neither beats the handbuilt descriptor: geometry-only accuracy is 0.566±0.013 (handbuilt), 0.571±0.007 (PointNet++), and 0.579±0.014 (OpenShape), and fused accuracy is 0.591±0.004, 0.588±0.002, and 0.592±0.006 respectively (DINOv2-giant,휆 = 0.35; all differences within seed standard deviation). The handbuilt-versus-learned tie reproduces on HOPE: PointNet++ reaches geometry-only 0.929±0.007 and fused 0.867±0.004 against the handbuilt descriptor’s 0.920±0.006 and 0.872±0.002, differences below one point in both arms. That a large-scale-pretrained 3D foundation encoder only matches a 45-dimensional handbuilt descriptor indicates that encoder sophistication is not what drives the result; geometry quality is (consistent with R2). We retain the training-free descriptor, which is lightweight and needs no point-cloud corpus; a frozen 3D foundation encoder remains a drop-in alternative at equal accuracy. Deployment cost Onboarding a new class is inexpensive. Reconstructing one object’s CAD-free 3DGS source from its scan takes on the order of 2.4–2.8 minutes per object on a single GPU and 0.14–0.17 GB of storage, across C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 13 of 18 A CAD-Free 3D Shape Prior for Object Recognition Table 3 Frozen image-encoder comparison (T-LESS held-out test set, 휆 = 0.35; mean±std over 5 seeds, image-only deterministic). Sorted by fused accuracy; DINOv2-giant is the encoder used in all other experiments. † 푛=4 (one seed failed). AIMv2 is omitted: its features were incompatible with the prototype pipeline (chance-level accuracy). Encoders are the public checkpoints facebook/dinov2-giant, google/siglip2-giant-opt-patch16-384, google/siglip2-so400m-patch16-384, google/siglip-so400m-patch14-384, google/siglip-large-patch16-256, laion/CLIP-ViT-H-14-laion2B-s32B-b79K, laion/CLIP-ViT-bigG-14-laion2B-39B-b160k, openai/clip-vit-large-patch14, facebook/metaclip-h14-fullcc2.5b, facebook/dinov3-vith16plus- pretrain-lvd1689m, facebook/dinov3-vitl16-pretrain-lvd1689m, facebook/dinov2-with-registers-giant, facebook/dinov2-with-registers-large, and facebook/ijepa_vith14_1k. EncoderFamilyImage-onlyGeomFused SigLIP2-g † image–text 0.561±0.000 0.611±0.010 0.629±0.004 MetaCLIP-Himage–text 0.559±0.000 0.589±0.010 0.613±0.006 SigLIP-Limage–text 0.545±0.000 0.555±0.002 0.592±0.005 DINOv2-giantSSL ViT0.560±0.000 0.566±0.013 0.591±0.004 SigLIP2-so400m image–text 0.531±0.000 0.553±0.013 0.586±0.007 OpenCLIP-Himage–text 0.546±0.000 0.516±0.016 0.578±0.011 DINOv3-LSSL ViT0.529±0.000 0.557±0.010 0.569±0.005 DINOv2-reg-gSSL ViT0.522±0.000 0.546±0.011 0.559±0.006 OpenCLIP-bigG image–text 0.548±0.000 0.477±0.027 0.556±0.012 SigLIP-so400m image–text 0.556±0.000 0.490±0.019 0.553±0.010 DINOv3-H+SSL ViT0.493±0.000 0.528±0.025 0.550±0.006 DINOv2-reg-LSSL ViT0.478±0.000 0.530±0.016 0.528±0.005 CLIP-Limage–text 0.316±0.000 0.336±0.034 0.361±0.019 I-JEPA-HJEPA0.118±0.000 0.134±0.021 0.152±0.020 Table 4 Frozen image-encoder comparison on HOPE (held-out test set, 휆 = 0.35; mean±std over 5 seeds, image-only deterministic). Same 14 encoders and protocol as Table 3, sorted by fused accuracy. HOPE objects are textured, so image-only accuracy is high; the geometry fusion still helps nearly every backbone, most where the backbone is weakest. DINOv2-giant (used elsewhere) is mid-pack here. EncoderFamilyImage-onlyGeomFused OpenCLIP-Himage–text 0.956±0.000 0.956±0.003 0.966±0.001 DINOv3-H+SSL ViT0.951±0.000 0.968±0.002 0.963±0.002 OpenCLIP-bigG image–text 0.950±0.000 0.943±0.006 0.956±0.002 SigLIP2-gimage–text 0.923±0.000 0.975±0.003 0.949±0.002 DINOv3-LSSL ViT0.934±0.000 0.963±0.008 0.947±0.001 SigLIP-so400m image–text 0.916±0.000 0.957±0.006 0.938±0.002 SigLIP2-so400m image–text 0.901±0.000 0.966±0.005 0.938±0.002 SigLIP-Limage–text 0.908±0.000 0.948±0.006 0.928±0.002 MetaCLIP-Himage–text 0.879±0.000 0.939±0.004 0.917±0.003 CLIP-Limage–text 0.869±0.000 0.934±0.003 0.911±0.003 DINOv2-giantSSL ViT0.832±0.000 0.920±0.006 0.872±0.002 DINOv2-reg-gSSL ViT0.839±0.000 0.932±0.003 0.868±0.003 DINOv2-reg-LSSL ViT0.796±0.000 0.853±0.014 0.836±0.005 I-JEPA-HJEPA0.155±0.000 0.349±0.023 0.273±0.013 C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 14 of 18 A CAD-Free 3D Shape Prior for Object Recognition onboarding coverages from 32 to 160 views (Table 5); the cost is dominated by 3DGS training (134–156 s), while frame preparation and point-cloud export together add only a few seconds. For the 160-view setting used in our main experiments this is about 2.8 minutes and 0.17 GB per object. The geometry prototype itself is a 45-dimensional vector, so the shape prior adds negligible storage and matching cost on top of the frozen image features. Cost grows only weakly with onboarding coverage (≈ +0.4 minutes per object from 32 to 160 views): denser scans buy robustness cheaply. Coverage refers to the azimuth-ring structure of the onboarding scan: each ring holds 32 views at a fixed elevation, so the single equatorial ring uses 32 views, three rings 96, and the full five-ring scan 160 (the main setting). We report this as a single-GPU (NVIDIA RTX 5000 Ada) engineering accounting, not a hardware-normalized benchmark. Table 5 CAD-free onboarding cost on T-LESS (28 objects, splatfacto-big, 20k iterations, single GPU). Reconstruction time and storage per object across onboarding coverages; the 160-view five-ring setting is the one used in the main experiments. Onboarding coverageViews/object Recon. time (min/obj) Storage (GB/obj) Single ring (equatorial)322.410.14 Three-ring962.660.15 Five-ring (main)1602.830.17 5. Discussion What this study establishes, and what it does not We state our claims only at the level the evidence supports. Three findings are each corroborated by several independent experiments. The prior is reconstruction-agnostic, so a short scan recovers the recognition value of CAD without CAD: RGB-D depth, 3DGS, and CAD geometry give comparable recognition, within seed-level variability (R2), and rendered pixels, unlike the geometry, add nothing (R5). The effect does not depend on any rendering-specific property of 3DGS. A second finding is complementarity: the prior helps most where appearance is least reliable, as shown by the occlusion stratification on both datasets (R4), the 14- backbone sweep in which the fusion gain grows as the backbone weakens (R6), the image-difficulty-controlled partial correlation (R3), and the feature-manifold analysis (Fig. 8). And the result is not driven by descriptor sophistication: a 45-dimensional handbuilt descriptor, an in-domain PointNet++, and a frozen Objaverse-pretrained 3D foundation encoder all perform within seed noise (R7). We deliberately make no stronger claims than these. We do not claim that fixed-weight fusion is the optimal configuration in every regime: on HOPE, where geometry dominates appearance, geometry alone is the ceiling and the untuned fusion sits below it (Fig. 10). We do not claim end-to-end detection-and-recognition, only recognition on labeled crops. Nor do we claim validation at large scale: our evidence spans 56 objects across two domains. The contribution is a carefully bounded mechanism, not a universal recipe; the rest of this section examines its boundaries. We initially expected the prior’s value to scale with raw shape distinctiveness, but within either dataset a purely geometric distinctiveness metric did not predict the per-class gain. The supported statement is the weaker, controlled one: geometry helps in proportion to its own per-class recognizability once image difficulty is held fixed (R3). The cross-dataset contrast (geometry matters more on HOPE than on T-LESS) is confounded by texture, shape, and overall difficulty: HOPE is easier overall, so it should be read as a trend, not a single-variable result. The fusion weight is fixed and shared across datasets by design; on HOPE a geometry-heavier weight would recover the geometry-only accuracy, but we prefer to avoid per-dataset tuning. We use 3DGS only as a practical way to obtain an object point cloud without CAD. Because the resulting geometry matches gold-standard CAD geometry (R2), any reconstruction that yields a comparable point cloud (e.g., multi-view stereo) would be expected to serve equally well; a systematic comparison of reconstruction methods remains open. Relatedly, our onboarding reconstructions use the calibrated object-centric poses supplied with each scan. Recovering such poses CAD-free in the wild, for example via structure-from-motion or video onboarding as in OnePose (Sun et al., 2022), is an orthogonal onboarding step; recognition depends on the resulting geometry, not its source (R2), so a comparable point cloud recovered that way should serve just as well. We also report recognition accuracy, not BOP detection or pose average precision; the work complements pose-benchmark methods but is not directly comparable with them. Our queries are cropped, labeled boxes, so the study isolates recognition from end-to-end detection-and-recognition. In deployment the prior would sit behind a CAD-free detector or class-agnostic C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 15 of 18 A CAD-Free 3D Shape Prior for Object Recognition segmenter (for example SAM-based proposals as in CNOS) that supplies the crops; integrating it into a full detect- then-recognize pipeline and measuring end-to-end accuracy remains future work. Confirming the prior at the scale of thousands of objects (e.g., large open-vocabulary or synthetic-object corpora), well beyond the 56 studied here, is an important next step. Finally, we match on a single global per-class feature prototype; intermediate patch-level descriptors, shown to carry positional information valuable under symmetry and texture loss (Örnek et al., 2024), are a complementary representation we leave to future work. 6. Conclusion A short object-centric scan recovers the recognition value of CAD without CAD. The geometry prior it yields nearly matches CAD geometry and pays off in proportion to how recognizable the object’s shape is. Fused with frozen foundation image features, it helps most where appearance fails: on geometrically recognizable but visually confusable objects and under partial occlusion. Onboarding needs nothing beyond the scan, and the prior is a compact per-class shape descriptor, so the approach offers a practical route to enrolling new parts on production lines where CAD is unavailable. Integrating it into a full detection-and-pose pipeline is left to future work. CRediT authorship contribution statement Chenxi Tao: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Seung-Kyum Choi: Supervision, Funding acquisition, Conceptualization, Writing – review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The T-LESS and HOPE datasets used in this study are publicly available through the BOP benchmark. Scan-derived reconstructions and analysis code are available from the corresponding author on reasonable request. Funding This research was funded by the Ministry of Trade, Industry and Energy (MOTIE), Korea, through the Global Industrial Technology Cooperation Center (GITCC) Program, supervised by the Korea Institute for Advancement of Technology (KIAT) (Task No. P24680172). The funding source had no involvement in study design; in the collection, analysis and interpretation of data; in the writing of the report; or in the decision to submit the article for publication. Acknowledgements The authors thank the GITCC Program (MOTIE/KIAT) for its support, and gratefully acknowledge the maintainers of the T-LESS and HOPE datasets and the BOP benchmark, and the open-source Nerfstudio and DINOv2 projects, whose resources made this study possible. References Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N., 2023. Self-supervised learning from images with a joint-embedding predictive architecture, in: CVPR. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P., 2022. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields, in: CVPR. Bortolon, M., Tsesmelis, T., James, S., Poiesi, F., Del Bue, A., 2024. 6DGS: 6D pose estimation from a single image and a 3D gaussian splatting model, in: European Conference on Computer Vision (ECCV). Cai, D., Heikkilä, J., Rahtu, E., 2022. OVE6D: Object viewpoint encoding for depth-based 6D object pose estimation, in: CVPR. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 16 of 18 A CAD-Free 3D Shape Prior for Object Recognition Cai, D., Heikkilä, J., Rahtu, E., 2025. GS-Pose: Generalizable segmentation-based 6D object pose estimation with 3D gaussian splatting, in: International Conference on 3D Vision (3DV), p. 1001–1011. Caraffa, A., Boscaini, D., Hamza, A., Poiesi, F., 2024. FreeZe: Training-free zero-shot 6D pose estimation with geometric and vision foundation models, in: ECCV. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A., 2021. Emerging properties in self-supervised vision transformers, in: ICCV. Clement, F., Shah, K., Pancholi, D., Lugo Bustillo, G., Cheng, I., 2024. Textureless object recognition: An edge-based approach. arXiv preprint arXiv:2403.06107 . Darcet, T., Oquab, M., Mairal, J., Bojanowski, P., 2024. Vision transformers need registers, in: International Conference on Learning Representations (ICLR). Denninger, M., Winkelbauer, D., Sundermeyer, M., Boerdijk, W., Knauer, M., Strobl, K.H., Humt, M., Triebel, R., 2023. BlenderProc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software 8, 4901. Drost, B., Ulrich, M., Bergmann, P., Härtinger, P., Steger, C., 2017. Introducing MVTec ITODD – a dataset for 3D object recognition in industry, in: ICCV Workshops. Dwibedi, D., Misra, I., Hebert, M., 2017. Cut, paste and learn: Surprisingly easy synthesis for instance detection, in: ICCV. Fini, E., et al., 2025. Multimodal autoregressive pre-training of large vision encoders (AIMv2), in: CVPR. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B., 2021. Simple copy-paste is a strong data augmentation method for instance segmentation, in: CVPR. Guédon, A., Lepetit, V., 2024. SuGaR: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R., 2022. Masked autoencoders are scalable vision learners, in: CVPR. Hinterstoisser, S., Lepetit, V., Ilic, S., Holzer, S., Bradski, G., Konolige, K., Navab, N., 2012. Model based training, detection and pose estimation of texture-less 3D objects in heavily cluttered scenes, in: Asian Conference on Computer Vision (ACCV). Hodaň, T., Haluza, P., Obdržálek, Š., Matas, J., Lourakis, M., Zabulis, X., 2017. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects, in: IEEE Winter Conference on Applications of Computer Vision (WACV). Hodaň, T., Sundermeyer, M., Drost, B., et al., 2020. BOP challenge 2020 on 6D object localization, in: ECCV Workshops. Hönig, P., Thalhammer, S., Weibel, J.B., Hirschmanner, M., Vincze, M., 2025. Shape-biased texture agnostic representations for improved textureless and metallic object detection and 6D pose estimation , 8795–8804. Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S., 2024. 2D gaussian splatting for geometrically accurate radiance fields, in: ACM SIGGRAPH. Jin, Y., Prasad, V., Jauhri, S., Franzius, M., Chalvatzaki, G., 2025. 6DOPE-GS: Online 6D object pose estimation using gaussian splatting, in: IEEE/CVF International Conference on Computer Vision (ICCV). Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42. Kirillov, A., Mintun, E., Ravi, N., et al., 2023. Segment anything, in: ICCV. Labbé, Y., Manuelli, L., Mousavian, A., Tyree, S., Birchfield, S., Tremblay, J., Carpentier, J., Aubry, M., Fox, D., Sivic, J., 2022. MegaPose: 6D pose estimation of novel objects via render & compare, in: Conference on Robot Learning (CoRL). Laurentini, A., 1994. The visual hull concept for silhouette-based image understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 16, 150–162. Lin, J., Liu, L., Lu, D., Jia, K., 2024. SAM-6D: Segment anything model meets zero-shot 6D object pose estimation, in: CVPR. Liu, M., Shi, R., Kuang, K., Zhu, Y., Li, X., Han, S., Cai, H., Porikli, F., Su, H., 2023. OpenShape: Scaling up 3D shape representation towards open-world understanding, in: NeurIPS. Liu, X., Wang, G., Li, C., Li, Y., Zhang, C., Huang, Z., Ji, X., 2024. GFreeDet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the BOP challenge 2024. arXiv preprint arXiv:2412.01552 . Ma, R., Zudaire, S., Li, Z., Zhang, C., 2025. IMD: A 6-DoF pose estimation benchmark for industrial metallic objects, in: International Conference on Robotics, Control and Automation Engineering (RCAE). Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R., 2020. NeRF: Representing scenes as neural radiance fields for view synthesis, in: European Conference on Computer Vision (ECCV). Müller, T., Evans, A., Schied, C., Keller, A., 2022. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41. Nguyen, V.N., Groueix, T., Ponimatkin, G., Lepetit, V., Hodan, T., 2023. CNOS: A strong baseline for CAD-based novel object segmentation, in: ICCV Workshops. Nguyen, V.N., Groueix, T., Salzmann, M., Lepetit, V., 2024. GigaPose: Fast and robust novel object pose estimation via one correspondence, in: CVPR. Nguyen, V.N., Tyree, S., Guo, A., Fourmy, M., Gouda, A., Lee, T., et al., 2025. BOP challenge 2024 on model-based and model-free 6D object pose estimation. arXiv preprint arXiv:2504.02812 . Oquab, M., Darcet, T., Moutakanni, T., et al., 2024. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research . Örnek, E.P., Labbé, Y., Tekin, B., Ma, L., Keskin, C., Forster, C., Hodan, T., 2024. FoundPose: Unseen object pose estimation with foundation features, in: ECCV. Osada, R., Funkhouser, T., Chazelle, B., Dobkin, D., 2002. Shape distributions. ACM Transactions on Graphics 21, 807–832. Qi, C.R., Su, H., Mo, K., Guibas, L.J., 2017a. PointNet: Deep learning on point sets for 3D classification and segmentation, in: CVPR. Qi, C.R., Yi, L., Su, H., Guibas, L.J., 2017b. PointNet++: Deep hierarchical feature learning on point sets in a metric space, in: NeurIPS. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 17 of 18 A CAD-Free 3D Shape Prior for Object Recognition Radford, A., Kim, J.W., Hallacy, C., et al., 2021. Learning transferable visual models from natural language supervision, in: International Conference on Machine Learning (ICML). Shugurov, I., Li, F., Busam, B., Ilic, S., 2022. OSOP: A multi-stage one shot object pose estimation framework, in: CVPR. Siméoni, O., et al., 2025. DINOv3. arXiv preprint arXiv:2508.10104 . Snell, J., Swersky, K., Zemel, R., 2017. Prototypical networks for few-shot learning, in: NeurIPS. Sun, J., Wang, Z., Zhang, S., He, X., Zhao, H., Zhang, G., Zhou, X., 2022. OnePose: One-shot object pose estimation without CAD models, in: CVPR. Tao, C., Choi, S.K., 2026. Prior availability in industrial visual sim-to-real: A review of CAD-guided and CAD-unavailable regimes. arXiv:2605.30581. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P., 2017. Domain randomization for transferring deep neural networks from simulation to the real world, in: IROS. Tschannen, M., et al., 2025. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786 . Tyree, S., Tremblay, J., To, T., Cheng, J., Mosier, T., Smith, J., Birchfield, S., 2022. 6-DoF pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark, in: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Wang, G., Manhardt, F., Tombari, F., Ji, X., 2021a. GDR-Net: Geometry-guided direct regression network for monocular 6D object pose estimation, in: CVPR. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W., 2021b. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction, in: NeurIPS. Wen, B., Yang, W., Kautz, J., Birchfield, S., 2024. FoundationPose: Unified 6D pose estimation and tracking of novel objects, in: CVPR. Xiang, Y., Schmidt, T., Narayanan, V., Fox, D., 2018. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes, in: Robotics: Science and Systems (RSS). Xiong, H., Muttukuru, S., Xiao, H., Upadhyay, R., Chari, P., Zhao, Y., Kadambi, A., 2025. SparseGS: Sparse view synthesis using 3D gaussian splatting, in: International Conference on 3D Vision (3DV). Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C., 2024. Demystifying CLIP data, in: ICLR. Xue, L., Gao, M., Xing, C., Martín-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S., 2023. ULIP: Learning a unified representation of language, images, and point clouds for 3D understanding, in: CVPR. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J., 2022. Point-BERT: Pre-training 3D point cloud transformers with masked point modeling, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L., 2023. Sigmoid loss for language image pre-training, in: ICCV. C. Tao and S.-K. Choi: Preprint submitted to ElsevierPage 18 of 18