Pith. sign in

REVIEW 2 major objections 4 minor 68 references

Zero-shot Inexact CAD Model Alignment from a Single Image

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper shows that a single image can align an inexact 3D CAD model to an object in 9 degrees of freedom without pose annotations, using geometry-aware features and NOC-based refinement.

desk verdict Strong method and a useful new benchmark, but the zero-shot claim is overstated because the metric depth estimator is fine-tuned on each target dataset; the headline margins likely depend on that adaptation. read the letter →

arxiv 2507.03292 v1 pith:PDYE6DQI submitted 2025-07-04 cs.CV

classification cs.CV
keywords single-view3Dalignment9-DoFposeestimationinexactCADmodelszero-shotgeneralizationfoundationfeaturesnormalizedobjectcoordinatesweaklysupervisedlearningcoarse-to-finerefinement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tackles the problem of aligning an inexact 3D CAD model to an object in a single photograph, estimating a 9-DoF pose (rotation, translation, and anisotropic scale) without any pose-annotated training data. The authors claim that a lightweight adapter trained only on synthetic renderings of CAD models can reshape generic foundation features so that object parts become geometry-aware, breaking the left-right symmetries that confuse off-the-shelf features. They combine this with a coarse-to-fine scheme: an initial pose from nearest-neighbor matching in the new feature space, refined by dense alignment in Normalized Object Coordinates, which is invariant to texture mismatch. On ScanNet25k the method is the only weakly supervised approach to beat the supervised ROCA baseline, and on the new SUN2CAD benchmark with 20 unseen categories it leads by a large margin. The significance is that CAD alignment could scale to novel object categories without expensive pose annotations.

What carries the argument

The load-bearing mechanism is the geometry-aware feature space produced by the adapter $E_\theta$, a two-layer MLP that maps DINOv2 features to adapted features, trained on multi-view CAD renderings with two losses: a NOC prediction loss ($\mathcal{L}_{\mathrm{NOC}}$) and a geometry-consistent triplet loss ($\mathcal{L}_{\mathrm{triplet}}$). The adapted features are fused with normalized DINOv2 features by a weight $\omega$ and used to build a feature voxel grid of the CAD model via back-projection of multi-view renderings. Coarse pose estimation then matches object-pixel features to voxel features and solves 3D-3D correspondences. Fine alignment optimizes the pose by rendering the model into a NOC map and minimizing an L1 NOC alignment loss plus silhouette and metric-depth losses, which makes the refinement texture-invariant and stable. The paper identifies the triplet loss as essential for separating symmetric parts (e.g., chair legs) that foundation features confuse, and the NOC-space comparison as essential for handling textureless or differently textured CAD models.

What would settle it

Run the open-source pipeline on ScanNet25k and SUN2CAD with an off-the-shelf metric depth estimator that is NOT fine-tuned on the target dataset, and compare mean alignment accuracy against the reported numbers; if accuracy does not drop substantially, the fine-tuned depth is not load-bearing, while a large drop would confirm the dependency.

Watch

Extended reading notes

Core claim

The paper's central claim is that zero-shot 9-DoF alignment of inexact CAD models is achievable from a single image using only synthetic, pose-free CAD renderings for supervision. The key discovery is that a small MLP feature adapter trained with a two-part self-supervised objective can turn DINOv2 foundation features into a geometry-aware space: a NOC prediction loss forces features to encode 3D location, and a triplet loss pulls together features of the same 3D part across views while pushing apart features of distant or symmetric parts. In this space, 2D image features and 3D model voxel features become directly comparable, so correspondences can be found by nearest-neighbor matching and a coarse pose solved with a robust pose solver using metric depth. The pose is then refined by differentiable rendering that compares, in NOC space, the predicted map of the input with the rendered map of the model, together with mask and depth losses. The authors report that this pipeline surpasses weakly supervised baselines by +4.3% mean alignment accuracy on ScanNet25k, becomes the only weakly supervised method to exceed supervised ROCA by +2.7%, and on the new SUN2CAD dataset of 20 unseen categories outperforms the supervised SPARC by +12.7%.

Load-bearing premise

The whole pipeline assumes the metric depth estimator, fine-tuned on the training images of each target dataset, predicts absolute scale accurately enough that lifting 2D features to 3D and the depth refinement loss stay reliable; if depth is biased, both coarse and fine pose estimates degrade and the zero-shot claim weakens.

Editorial extensions

If this is right

  • CAD retrieval-and-alignment systems can be trained for new object categories using only synthetic renderings, removing the need for category-specific pose annotations.
  • A pose estimated from a single image can be refined purely in normalized object coordinates, so textureless or differently textured CAD models no longer break dense alignment.
  • On the introduced SUN2CAD benchmark, the method achieves state-of-the-art generalization to 20 unseen categories without prior training on them, suggesting category-agnostic alignment is feasible.
  • The geometry-aware adapter could improve other tasks that build on foundation features, such as semantic correspondence or 6-DoF pose estimation, whenever part-level symmetry is a confound.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'zero-shot' label is qualified by the need to fine-tune the metric depth estimator on each target dataset; a truly zero-shot variant would have to rely on a depth model that generalizes without per-dataset fine-tuning.
  • The feature adapter is trained on only nine ShapeNet categories yet transfers to 20 unseen SUN2CAD categories, which suggests that the learned geometry-awareness is not category-specific and could improve with more diverse synthetic training data.
  • The dependence on a predicted metric depth map for both coarse lifting and fine depth loss means the method's accuracy is bounded by depth quality; combining it with a different depth source (e.g., multi-view stereo) is a natural next step the paper does not explore.
  • The SUN2CAD benchmark with manual refinement could be reused to quantify the relationship between CAD model inexactness (Chamfer distance) and pose accuracy, since the paper only studies this on 100 ScanNet images.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a weakly supervised 9-DoF alignment method for inexact CAD models from a single RGB image. It enhances DINOv2 features with a lightweight MLP adapter trained on synthetic renderings using a NOC regression loss and a geometry-consistent triplet loss, then uses nearest-neighbor matching in the fused feature space together with a metric-depth estimator to solve for a coarse pose. A dense alignment stage refines the pose by comparing the input's predicted NOC map, silhouette, and depth to differentiable renderings of the posed model. The method is evaluated on ScanNet25k, DiffCAD's split, and a new SUN2CAD dataset with 20 unseen categories, where it is claimed to be the only weakly supervised approach to surpass the supervised ROCA and to achieve large margins on unseen categories.

Significance. If the central claims hold, the paper makes a useful contribution: it demonstrates that a small feature adapter trained on synthetic renderings can produce geometry-aware features that transfer to unseen categories, and it provides a texture-invariant NOC-space refinement that improves over feature-metric alternatives. The new SUN2CAD benchmark addresses a real gap in evaluating generalization to novel categories, and the ablation study in Table 1 convincingly isolates the gains from fused features (+5.2%) and NOC refinement (+4.8%). The paper is generally well written and includes extensive appendices with additional ablations and qualitative results. The main caveat is that the reported margins depend on fine-tuning the metric depth estimator on each target dataset, which is not reflected in the 'zero-shot' characterization; quantifying that dependence would substantially strengthen the paper.

major comments (2)
  1. [Section 3.1.3, Eq. (5), Appendix 6.6] The 'zero-shot' claim is confounded by target-domain fine-tuning of the metric depth estimator. Appendix 6.6 states that DepthAnything is fine-tuned on the training images of ScanNet25k and SUN-RGBD, and the method uses metric depth both to lift 2D features to 3D in coarse alignment (Section 3.1.3) and as a depth loss in fine alignment (Eq. 5). Since the evaluation thresholds include absolute translation (20 cm) and scale (20%), any bias in the fine-tuned depth directly affects pass/fail, so the reported margins over ROCA (+2.7% instance-wise on ScanNet25k) and SPARC (+12.7% on SUN2CAD) are not attributable solely to the proposed feature/NOC pipeline. The paper already uses a non-fine-tuned depth estimator for the DiffCAD comparison in Section 4.2, but no analogous off-the-shelf-depth ablation is reported for the main Table 1 NMS metric or Table 2 SUN2CAD results. I request that the authors report these results with off-the-shelf DepthAnything and either qualify the 'zero-shot' statement to specify that only the feature adapter and NOC predictor are category-zero-shot, or show that depth fine-tuning has a negligible effect on the headline numbers.
  2. [Section 5 and Abstract] The conclusion states that the method is 'fully self-supervised, requiring only a lightweight adapter trained on a small number of ScanNet25K categories', and the abstract says the method works 'without prior training on' the 20 unseen categories. These statements are inaccurate because the metric depth estimator is fine-tuned on the training images of each target dataset (Appendix 6.6), which is a form of target-domain supervision. The SUN2CAD training images are drawn from SUN-RGBD scenes that contain the same object categories as the test set (though disjoint scenes), so 'without prior training on them' is misleading. Please correct these characterizations or add an explicit caveat that the zero-shot property applies to the pose-alignment components only, not to the depth backbone.
minor comments (4)
  1. [Section 4.1] The text says the method surpasses ROCA 'by {+2.7%,+1.6%}' for category-wise and instance-wise NMS scores, but Table 1 shows the opposite pairing: ROCA Avg Cat is 21.5 vs Ours 23.1 (+1.6) and Avg Inst is 27.4 vs 30.1 (+2.7). Please correct the order of the braces.
  2. [Abstract and Section 1] The abstract reports '+4.3% mean alignment accuracy' over SOTA weakly supervised baselines, while the introduction reports '+4.2%'. Table 1 shows +4.2% instance-wise vs DiffCAD (Err) on DiffCAD's split, so the +4.3% does not exactly match any value in the table. Please reconcile these numbers.
  3. [Section 4 and Appendix 6.6] The main text should explicitly state at the start of Section 4 that the depth estimator is fine-tuned on the target dataset for the ScanNet25k and SUN2CAD comparisons, and that it is not fine-tuned for the DiffCAD-split comparison. Currently this is only disclosed in the appendix, which is easy to miss.
  4. [Section 3 and Appendix 9] The method description says 'we use ROCA's retrieval system [18]', but for the new SUN2CAD dataset the CAD models are manually selected from Objaverse/ShapeNet. Please clarify in the main text that the retrieval system is used only for ScanNet25k, while SUN2CAD uses human-provided model selection.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; disclosed per-dataset depth fine-tuning is an attribution caveat, not a circular reduction.

full rationale

The core pipeline is self-contained. The feature adapter Eθ is trained on synthetic ShapeNet renderings with a NOC-regression loss (Eq. 1) and a triplet loss (Eq. 2); the decoder is discarded and no pose annotations are used. At inference, coarse alignment derives 2D–3D correspondences by nearest-neighbor matching in the fused DINOv2/adapter feature space and solves the pose with RANSAC, while fine alignment optimizes the NOC, silhouette, and depth losses (Eqs. 3–5) against renderings of the given CAD model. The NOC map used in refinement is a nearest-neighbor lookup in the same feature space; this is a standard correspondence-based refinement, not a self-fulfilling fit to the evaluation labels, because final accuracy is measured against external Scan2CAD and SUN-RGBD pose annotations. SUN2CAD's 20 categories are disjoint from the 9 ShapeNet categories used for adapter training, so the generalization claim is tested on genuinely held-out classes. The only caveat, disclosed in Appendix 6.6, is that DepthAnything is fine-tuned on the training images of ScanNet25k and SUN-RGBD before the main comparisons; this is target-domain adaptation of an auxiliary component and affects attribution of the headline margins, but it is not a circular step because the depth estimator is not fitted to pose predictions and the paper explicitly reports a non-fine-tuned depth variant for the DiffCAD comparison. No load-bearing self-citation or imported uniqueness theorem appears; references to ScanNet, Scan2CAD, ROCA, and Vid2CAD are standard external datasets, baselines, and metrics.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

Central claim rests on a large set of domain assumptions, mostly standard for this line of work. The most consequential is that a per-dataset fine-tuned metric depth estimator is available; the paper is transparent about this but it limits the zero-shot generality. No new physical entities are introduced, and the free parameters are conventional hyperparameters chosen on validation splits.

free parameters (10)
  • beta (adapter loss weight) = 0.1
    Balances NOC prediction and triplet losses; chosen by validation on the ScanNet25k validation set (Figure 6 in the paper).
  • omega (feature fusion weight) = 0.5
    Weights the fusion of DINOv2 and adapted features; tuned on validation NOC error (Figure 6).
  • tau_plus_dist = 0.02
    Positive sample distance threshold for triplet loss; tuned on validation (Appendix 12.5).
  • tau_minus_dist = 0.4
    Negative sample minimum 3D distance for triplet loss; tuned on validation (Appendix 12.5).
  • tau_minus_feat = 0.75
    Negative sample feature similarity threshold for triplet loss; tuned on validation (Appendix 12.5).
  • alpha (triplet margin) = 0.5
    Margin for the triplet loss; part of the standard triplet formulation.
  • lambda_NOC_A = 0.33
    Weight for the NOC alignment loss in dense refinement; tuned on validation (Appendix 12.6).
  • lambda_m = 3.0
    Weight for the silhouette loss in dense refinement; tuned on validation (Appendix 12.6).
  • lambda_d = 0.27
    Weight for the depth loss in dense refinement; tuned on validation (Appendix 12.6).
  • NOC_error_filter_threshold = 0.20
    Filters out diffusion-generated training images whose predicted NOC error exceeds this threshold (Appendix 12.2).
assumptions (6)
  • domain assumption DINOv2 features provide a strong semantic prior for zero-shot visual correspondence.
    The entire method builds on DINOv2 as the base feature extractor (Section 3.1.1).
  • domain assumption Normalized Object Coordinates (NOC), derived from front-aligned CAD models, form a consistent canonical space for alignment.
    The NOC prediction loss and the NOC-space refinement both assume this representation (Sections 3.1.1 and 3.2).
  • domain assumption A feature adapter trained on nine ShapeNet categories transfers to 20 unseen SUN2CAD categories.
    The zero-shot claim depends on this transfer (Section 4.3 and Appendix 10.4).
  • domain assumption The per-dataset fine-tuned DepthAnything estimator provides accurate metric depth.
    Metric depth is used for 3D-3D pose solving and as the depth loss in refinement (Sections 3.1.3 and 3.2, Implementation details).
  • domain assumption SAM and Grounded-SAM produce accurate object masks.
    Masks restrict feature matching to the object and drive the silhouette loss (Section 3.1.3, Appendix 6.5).
  • domain assumption The nearest-neighbor matching in the feature voxel grid is robust to synthetic-to-real domain shift.
    This is the central inference mechanism, and the paper argues it is more robust than direct NOC regression (Sections 3.1.3 and 4.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-shot Inexact CAD Model Alignment from a Single Image." pith.science (2026). https://pith.science/paper/PDYE6DQI

@misc{pith2026250703292,
  author       = {Pith},
  title        = {Pith review of: Zero-shot Inexact CAD Model Alignment from a Single Image},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDYE6DQI}},
  note         = {Machine review of arXiv:2507.03292}
}
read the original abstract

One practical approach to infer 3D scene structure from a single image is to retrieve a closely matching 3D model from a database and align it with the object in the image. Existing methods rely on supervised training with images and pose annotations, which limits them to a narrow set of object categories. To address this, we propose a weakly supervised 9-DoF alignment method for inexact 3D models that requires no pose annotations and generalizes to unseen categories. Our approach derives a novel feature space based on foundation features that ensure multi-view consistency and overcome symmetry ambiguities inherent in foundation features using a self-supervised triplet loss. Additionally, we introduce a texture-invariant pose refinement technique that performs dense alignment in normalized object coordinates, estimated through the enhanced feature space. We conduct extensive evaluations on the real-world ScanNet25k dataset, where our method outperforms SOTA weakly supervised baselines by +4.3% mean alignment accuracy and is the only weakly supervised approach to surpass the supervised ROCA by +2.7%. To assess generalization, we introduce SUN2CAD, a real-world test set with 20 novel object categories, where our method achieves SOTA results without prior training on them.

Figures

Figures reproduced from arXiv: 2507.03292 by the authors.

Figure 1
Figure 1. We estimate the 9-DoF pose of an inexact CAD model to align it with a target object’s pose in a 2D image without training on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Method Overview. From an image-3D model pair, we first construct 2D and 3D feature grids from DINOv2 and our geometry￾aware adapter, trained with 3D self-supervision, then use nearest-neighbor matching to establish correspondences for initial pose solving. Finally, we refine the pose through dense alignment between the predicted NOC map and the 3D model’s rendered NOC map. map [48] for each CAD rendering, where each… view at source ↗
Figure 3
Figure 3. Visualization of our learned space. DINOv2 and our geometry-aware features are dimensionally reduced with PCA and color-coded by NOC (see Appendix 6.7). Nearby object parts are more clearly separated in our features compared to DINOv2. which lifts the 2D coordinates of each feature at (u, v) in the feature map to 3D (u, v, d). We then convert these to stan￾dard (x, y, z) via back-projection and estimate pose from 3D… view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: Qualitative results in ScanNet25k dataset. 4.3. Comparison on SUN2CAD Dataset To assess zero-shot capability on unseen categories, we introduce SUN2CAD, a new inexact 9-DoF test set that extends beyond the nine categories in ScanNet25k [11]. SUN2CAD comprises 550 sampl…
Figure 6
Figure 6. Figure 6: Ablation studies on feature adapter losses (upper) and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: Qualitative results in SUN2CAD dataset. 4.5. Ablation Study on Hyperparameters Feature adapter loss study. The impact of β for balancing feature adapter loss in Eq 1 and Eq 2 is shown in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Rendered templates used in a feature voxel grid and the geometry-aware feature adapter. Uni-ControlNet Template Conditions Augmented Templates Canny Edge Depth Sketch … … … … … … #1 #2 #3 #4 #5 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Augmented templates used in a feature voxel grid and the geometry-aware feature adapter. Annotated red boxes refer to invalid generated images with a wrong viewpoint shifted from their original renderings on the top row. 8. Implementation Details of Competitors and Abl…
Figure 9
Figure 9. Figure 9: Qualitative results in NOC map prediction on unseen ScanNet25k categories. Our method reliably generates more accurate NOC maps than DINOv2 and is robust against unseen categories, unlike NOC-R, which excels in standard settings but struggles in unfamiliar categories. …
Figure 10
Figure 10. Figure 10: Rotation accuracy under varying CAD model inexact [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results in NOC prediction on Scan￾Net25k dataset. Our solution provides better smoothness and cor￾rectness than DINOv2, while NOC-S learns to output a smooth NOC map, multiple samples show shifted or invalid NOC range compared to ground-truth NOC maps. ate…
Figure 13
Figure 13. Figure 13: Qualitative results of dense image-based alignment optimization, shown before (coarse alignment) and after fine pose optimization. Architecture NOC error ± SE ↓ AE 0.2485 ± 0.0024 MLP 0.2263 ± 0.0023 ViT 0.2454 ± 0.0020 [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Ablation study in Triplet loss hyperparameters τ + dist, τ − dist, and τ − feat on NOC error metric. 0.5 1 3 4 5 7 m' 0.5 1 3 4 5 6 10 d' 27.44 27.70 27.12 26.44 26.18 25.96 29.28 29.17 28.91 29.27 27.81 27.24 29.23 30.38 30.59 30.28 29.75 29.64 29.35 30.75 31.33 30.7…
Figure 16
Figure 16. Figure 16: Failure cases; (A) Poor input image quality; (B) incor￾rect input mask, causing two objects’ features to blend together (C, D) for CAD alignment; (E) lack of edge cues for inferring object depth; (F, G) alignment ambiguity; (H) Thin object. and bottom, the thin struct…
Figure 17
Figure 17. Figure 17: Random test samples from ScanNet25k. We compare entire-scene CAD alignment against 9-DoF supervised methods (SPARC [26] and ROCA [18]), 9-DoF weakly supervised methods (DiffCAD [16]), and 6-DoF weakly supervised baselines (Foundation￾Pose [49]) [PITH_FULL_IMAGE:figur…
Figure 18
Figure 18. Figure 18: Random test samples from ScanNet25k. We compare entire-scene CAD alignment against 9-DoF supervised methods (SPARC [26] and ROCA [18]), 9-DoF weakly supervised methods (DiffCAD [16]), and 6-DoF weakly supervised baselines (Foundation￾Pose [49]) [PITH_FULL_IMAGE:figur…
Figure 19
Figure 19. Figure 19: Random test samples from ScanNet25k. We compare entire-scene CAD alignment against 9-DoF supervised methods (SPARC [26] and ROCA [18]), 9-DoF weakly supervised methods (DiffCAD [16]), and 6-DoF weakly supervised baselines (Foundation￾Pose [49]) [PITH_FULL_IMAGE:figur…
Figure 20
Figure 20. Figure 20: Random test samples from SUN2CAD comparing 9-DoF pose predictions with SPARC [26] and DINOv2, and 6-DoF predictions with FoundationPose [49] [PITH_FULL_IMAGE:figures/full_fig_p026_20.png]
Figure 21
Figure 21. Figure 21: Random test samples from SUN2CAD comparing 9-DoF pose predictions with SPARC [26] and DINOv2, and 6-DoF predictions with FoundationPose [49] [PITH_FULL_IMAGE:figures/full_fig_p027_21.png]
Figure 22
Figure 22. Figure 22: Random test samples from SUN2CAD comparing 9-DoF pose predictions with SPARC [26] and DINOv2, and 6-DoF predictions with FoundationPose [49] [PITH_FULL_IMAGE:figures/full_fig_p028_22.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 54 canonical work pages

  1. [1]

    Zs6d: Zero-shot 6d object pose estimation using vision transformers

    Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463–469. IEEE, 2024. 2

  2. [2]

    Scan2cad: Learning cad model alignment in rgb-d scans

    Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. InProceed- ings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614–2623, 2019. 1, 6, 4, 8

  3. [3]

    Depth pro: Sharp monocular metric depth in less than a second

    Alexey Bochkovskiy, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. InThe Thirteenth International Conference on Learning Representations, 2025. 2

  4. [4]

    Learning 6d object pose estimation using 3d object coordinates

    Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536–551. Springer, 2014. 2, 5

  5. [5]

    Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models.ECCV, 2024

    Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models.ECCV, 2024. 2, 3, 5

  6. [6]

    Emerg- ing properties in self-supervised vision transformers.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630–9640, 2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv’e J’egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630–9640, 2021. 3

  7. [7]

    Chang, Thomas A

    Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository.ArXiv, abs/1512.03012, 2015. 2, 3, 6, 1, 4

  8. [8]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. InProceedings of the 37th Inter- national Conference on Machine Learning. JMLR.org, 2020. 3

Show all 68 references
  1. [9]

    Open-vocabulary object 6d pose estimation

    Jaime Corsetti, Davide Boscaini, Changjae Oh, Andrea Cav- allaro, and Fabio Poiesi. Open-vocabulary object 6d pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18071– 18080, 2024. 2

  2. [10]

    Indoor semantic segmentation using depth in- formation

    Camille Couprie, Cl ´ement Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth in- formation. InFirst International Conference on Learning Representations (ICLR 2013), pages 1–8, 2013. 2

  3. [11]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas A

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes.2017 IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 2432–2443, 2017. 1, 2, 6, 7, 3, 4, 5

  4. [12]

    Obja- verse: A universe of annotated 3d objects.arXiv preprint arXiv:2212.08051, 2022

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Obja- verse: A universe of annotated 3d objects.arXiv preprint arXiv:2212.08051, 2022. 4

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

  6. [14]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981. 5

  7. [15]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Bin- qiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 10933–10942,

  8. [16]

    Diffcad: Weakly-supervised probabilistic cad model retrieval and alignment from an rgb image.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    Daoyi Gao, D ´avid Rozenberszki, Stefan Leutenegger, and Angela Dai. Diffcad: Weakly-supervised probabilistic cad model retrieval and alignment from an rgb image.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024. 1, 2, 3, 5, 6, 7, 4, 8, 11, 12, 13, 14

  9. [17]

    Complete solution classification for the perspective-three-point problem.IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943,

    Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem.IEEE transactions on pattern analysis and machine intelligence, 25(8):930–943,

  10. [18]

    Roca: Ro- bust cad model retrieval and alignment from a single image

    Can G ¨umeli, Angela Dai, and Matthias Nießner. Roca: Ro- bust cad model retrieval and alignment from a single image. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4022–4031, 2022. 1, 2, 3, 5, 6, 7, 4, 12, 13, 14

  11. [19]

    LVIS: A dataset for large vocabulary instance segmentation

    Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4

  12. [20]

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geomet- ric foundation model for zero-shot metric depth and surface normal estimation.IEEE Transactions on Pattern Analysi...

  13. [21]

    Matchu: Matching unseen objects for 6d pose estimation from rgb-d images

    Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slo- bodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095–10105, 2024. 2

  14. [22]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B. Girshick. Segment anything.2023 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), p...

  15. [23]

    Mask2cad: 3d shape prediction by learning to segment and retrieve

    Weicheng Kuo, Anelia Angelova, Tsung-Yi Lin, and Angela Dai. Mask2cad: 3d shape prediction by learning to segment and retrieve. InComputer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part III 16, pages 260–277. Springer, 2020. 1, 3, 6

  16. [24]

    Patch2cad: Patchwise embedding learning for in-the- wild shape retrieval from a single image

    Weicheng Kuo, Anelia Angelova, Tsung-Yi Lin, and Angela Dai. Patch2cad: Patchwise embedding learning for in-the- wild shape retrieval from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12589–12599, 2021. 1, 3

  17. [25]

    Megapose: 6d pose estimation of novel objects via render and compare

    Yann Labb ´e, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpen- tier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render and compare. InProceedings of The 6th Conference on Rob...

  18. [26]

    Langer, G

    F. Langer, G. Bae, I. Budvytis, and R. Cipolla. Sparc: Sparse render-and-compare for cad model alignment in a single rgb image. InProc. British Machine Vision Conference, London,

  19. [27]

    Mrc- net: 6-dof pose estimation with multiscale residual correla- tion

    Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrc- net: 6-dof pose estimation with multiscale residual correla- tion. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 10476–10486. IEEE,

  20. [28]

    Deep learning-based object pose estimation: A comprehensive survey.arXiv preprint arXiv:2405.07801,

    Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey.arXiv preprint arXiv:2405.07801,

  21. [29]

    Soft ras- terizer: A differentiable renderer for image-based 3d reason- ing.The IEEE International Conference on Computer Vision (ICCV), 2019

    Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft ras- terizer: A differentiable renderer for image-based 3d reason- ing.The IEEE International Conference on Computer Vision (ICCV), 2019. 5

  22. [30]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detec- tion.ArXiv, abs/2303.05499, 2023

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chun yue Li, Jianwei Yang, Hang Su, Jun-Juan Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detec- tion.ArXiv, abs/2303.05499, 2023. 3

  23. [31]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2017. 6, 1, 9

  24. [32]

    Generating images with 3d annotations using diffusion models

    Wufei Ma, Qihao Liu, Jiahao Wang, Angtian Wang, Xiaod- ing Yuan, Yi Zhang, Zihao Xiao, Guofeng Zhang, Beijia Lu, Ruxiao Duan, et al. Generating images with 3d annotations using diffusion models. InThe Twelfth International Confer- ence on Learning Representations, 2024. 2

  25. [33]

    Vid2cad: Cad model alignment using multi- view constraints from videos.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1320–1327, 2020

    Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multi- view constraints from videos.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1320–1327, 2020. 6, 8

  26. [34]

    Im- proving semantic correspondence with viewpoint-guided spherical maps

    Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Im- proving semantic correspondence with viewpoint-guided spherical maps. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521– 19530, 2024. 2, 3

  27. [35]

    Gigapose: Fast and robust novel object pose estimation via one correspondence

    Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2

  28. [36]

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael ...

  29. [37]

    Found- pose: Unseen object pose estimation with foundation fea- tures.European Conference on Computer Vision (ECCV),

    Evin Pınar ¨Ornek, Yann Labb ´e, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tom ´aˇs Hoda ˇn. Found- pose: Unseen object pose estimation with foundation fea- tures.European Conference on Computer Vision (ECCV),

  30. [38]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InThe Twelfth Interna- tional Conference on Learning Representations, 2024. 3

  31. [39]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conf...

  32. [40]

    Accelerating 3d deep learning with pytorch3d

    Nikhila Ravi, Jeremy Reizenstein, David Novotny, Tay- lor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5, 6, 3

  33. [41]

    Grounded sam: Assembling open-world models for diverse visual tasks,

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kun- chang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks,

  34. [42]

    Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer

    Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, 2021. 3, 2, 7, 8

  35. [43]

    Facenet: A unified embedding for face recognition and clus- tering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 815–823, 2015. 2, 4

  36. [44]

    Dpodv2: Dense correspondence-based 6 dof pose estima- tion.IEEE transactions on pattern analysis and machine intelligence, 44(11):7417–7435, 2021

    Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estima- tion.IEEE transactions on pattern analysis and machine intelligence, 44(11):7417–7435, 2021. 2

  37. [45]

    Sun rgb-d: A rgb-d scene understanding benchmark suite

    Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567–576, 2015. 7, 1, 3

  38. [46]

    Onepose: One-shot object pose estimation without cad mod- els

    Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad mod- els. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 6825–6834,

  39. [47]

    An eigendecomposition approach to weighted graph matching problems.IEEE transactions on pattern analysis and machine intelligence, 10(5):695–703,

    Shinji Umeyama. An eigendecomposition approach to weighted graph matching problems.IEEE transactions on pattern analysis and machine intelligence, 10(5):695–703,

  40. [48]

    Normalized object coordinate space for category-level 6d object pose and size estimation

    He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26...

  41. [49]

    Foundationpose: Unified 6d pose estimation and tracking of novel objects

    Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868– 17879, 2024. 1, 2, 3, 5, 6, 7, 4, 8, 12, 13, 14, 15, 16, 17

  42. [50]

    Depth anything: Unleashing the power of large-scale unlabeled data.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371–10381, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371–10381, 2024. 3, 4, 6, 1, 2

  43. [51]

    Depth any- thing v2.NeurIPS, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.NeurIPS, 2024. 3

  44. [52]

    A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Pola- nia Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. InThirty- seventh Conference on Neural Information Processing Sy...

  45. [53]

    Telling left from right: Identifying geometry-aware semantic corre- spondence.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3076–3085, 2023

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic corre- spondence.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3076–3085, 2023. 2, 3

  46. [54]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 2

  47. [55]

    Uni-controlnet: All-in-one control to text-to-image diffusion models.Advances in Neural Information Processing Sys- tems, 36, 2024

    Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models.Advances in Neural Information Processing Sys- tems, 36, 2024. 3, 2 Zero-shot Inexact CAD Model Alignment fro...

  48. [57]

    Additional Implementation Details of Our method 6.1. Training details and hyperparameters For training the feature adapter, we use the AdamW opti- mizer [31] with a constant learning rate of3e −4, a batch size of 140, andτ + dist = 0.02,τ − dist = 0.4,τ − feat = 0.75,α= 0.5, β...

  49. [58]

    We fol- low DST3D [32] to generate realistic renderings with natu- ral backgrounds, using Stable Diffusion (SD) [42]

    Implementation Details of Rendered Image Augmentation Renderings often have limited texture and solid color back- grounds, creating a domain gap with real images. We fol- low DST3D [32] to generate realistic renderings with natu- ral backgrounds, using Stable Diffusion (SD) [4...

  50. [59]

    For ROCA, scores are reported as provided in their paper

    Implementation Details of Competitors and Ablation Choices In this section, we provide implementation details for evalu- ating DiffCAD, FoundationPose, and SPARC in our setting, where scores are computed using their official code. For ROCA, scores are reported as provided in t...

  51. [60]

    The primary challenge in aligning images with 3D mod- els is the ambiguity in z-axis translation and scaling, which persists even with manual alignment

    Additional Details on SUN2CAD Dataset To evaluate our alignment method on unseen or less com- mon object categories, we establish a new inexact match 9-DoF pose alignment test set spanning 20 categories with 550 images. The primary challenge in aligning images with 3D mod- els...

  52. [61]

    Additional Results on Experiment 4.1 We present a detailed comparison of translation, rotation, and scaling accuracies across competitors on the Scan- Net25k dataset in Table 4

    Additional Experimentsal Results 10.1. Additional Results on Experiment 4.1 We present a detailed comparison of translation, rotation, and scaling accuracies across competitors on the Scan- Net25k dataset in Table 4. All accuracies are computed using the same thresholds define...

  53. [62]

    Additional Qualitative Results 11.1. Qualitative Results in NOC Map Prediction Figure 11 and Figure 12 show the NOC prediction results of our method and other NOC predictor baselines, NOC- S and DINOv2, on ScanNet25k [11] and SUN2CAD, re- Group Method basket bicycle blender br...

  54. [63]

    Study on the Architecture of Feature Adapter We study the best network architecture for our feature adapter

    Additional Ablation Studies 12.1. Study on the Architecture of Feature Adapter We study the best network architecture for our feature adapter. We choose 3 choices: Autoencoder (AE), ViT [13] layers, and MLP layers. AE represents CNNs capable of decoding feature maps into pixel...

  55. [64]

    (for 9D) + GT depth 23.3 25.7 34.51.9 3.1 45.4 20.4 20.4 20.3 21.7 28.4 Ours 16.7 18.6 22.8 12.7 9.2 49.3 24.1 38.116.5 23.1 30.1 6D FoundationPose [49] 22.5 21.4 37.5 6.1 5.8 44.5 30.4 29.2 27.1 24.9 31.1

  56. [65]

    Our method still surpasses [49] + GT depth in average alignment accuracies

    + GT depth 25.8 27.1 47.09.0 7.3 46.7 32.5 38.1 33.1 29.6 35.0 Ours(for 6D) 20.8 25.7 27.6 19.8 22.7 56.1 51.8 45.120.1 32.2 38.0 Table 10.Comparison in NMS accuracy [33] on ScanNet25k [2] against FoundationPose [49].We additionally provide GT depth maps to FoundationPose to m...

  57. [66]

    Our method still surpasses [49] + GT depth in average alignment accuracies

    + GT depth28.650.014.3100.07.7 10.5 11.153.331.875.047.727.116.7 9.8 0.0 0.0 31.90.0 55.633.3 30.2 29.5 Ours(for 6D) 42.950.014.350.0 23.136.816.720.039.450.0 30.3 20.361.127.225.00.0 13.878.644.466.6 35.5 30.7 Table 11.Comparison in Single-view accuracy [16] on SUN2CAD agains...

  58. [67]

    Failure Cases Figure 16 illustrates failure cases. Poor input image quality, such as small objects (A), degrades DINOv2 features, mak- ing it difficult to distinguish object parts correctly, which can lead to incorrect pose estimation. Failures in dependencies, such as depth p...

  59. [68]

    However, its accuracy may be insufficient for safety-critical tasks like autonomous driving

    Societal Impacts Our work on 9-DoF pose estimation benefits real-world ap- plications in synthetic environments such as VR and gam- ing, where safety is not a concern. However, its accuracy may be insufficient for safety-critical tasks like autonomous driving. Reliance on pred...

  60. [2022]

    1, 2, 3, 5, 6, 7, 4, 12, 13, 14, 15, 16, 17

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.