Universal Features Guided Zero-Shot Category-Level Object Pose Estimation

Chenyu Meng; Cuixia Ma; Heng Li; Hongan Wang; Jian Cheng; Ping Tan; Wentian Qu; Xiaoming Deng; Xiao Zhou

REVIEW 3 major objections 5 minor 14 references

Universal Features Guided Zero-Shot Category-Level Object Pose Estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Pre-trained features unlock 6-DOF pose for unseen object categories.

desk verdict Solid zero-shot pose paper with strong benchmark numbers, but the 'zero-shot' claim is only as good as the hand-selected reference mesh and the missing error bars. read the letter →

arxiv 2501.02831 v1 pith:XGBWQJJV submitted 2025-01-06 cs.CV

Wentian Qu , Chenyu Meng , Heng Li , Jian Cheng , Cuixia Ma , Hongan Wang , Xiao Zhou , Xiaoming Deng

show 1 more author

Ping Tan

This is my paper · ORCID

classification cs.CV

keywords zero-shotposeestimationcategory-level6-DOFuniversalfeaturesDINOv2StableDiffusionDGCNNRGB-D

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 proposes a zero-shot method for category-level 6-DOF object pose estimation that works on object categories never seen during training. Instead of learning category-specific shape priors, it extracts pre-trained 'universal' features from both the RGB image (DINOv2 plus Stable Diffusion) and the depth point cloud (DGCNN) to establish correspondences between the observed object and a generic reference mesh of the same category. A coarse stage finds sparse 2D feature correspondences, solves for the 6-DOF pose with least squares, and iterates re-rendering to recover from large pose gaps; a refinement stage then jointly optimizes pose and reference shape using dense 3D feature alignment. The authors report that this pipeline outperforms prior supervised, self-supervised, and zero-shot methods on the REAL275 and Wild6D benchmarks for unseen categories, and that the gains come from the iterative coarse stage and the 3D universal alignment loss.

What carries the argument

The machinery is a cascade of universal-feature correspondences. 2D universal features are cosine-similarity descriptors from DINOv2 and Stable Diffusion overlaid patch-wise; the cyclical-distance step keeps the top-k mutually consistent matches. Umeyama with RANSAC lifts those 2D matches to a 3D similarity transform, and rendering the reference mesh under the current pose re-feeds the correspondences until convergence. The refinement stage introduces the universal alignment loss $L_g$, which computes the cosine similarity of DGCNN 3D features between reference and target point clouds and drives the positions of high-confidence pairs together, alongside mask and Chamfer losses and a regularization that keeps the deformed reference near its initial shape and pose. The whole cascade carries the argument: no learned pose head or category prior is used.

What would settle it

On a test set spanning six categories (REAL275), replace the reference mesh with one deliberately far from the target instances, such as a handle-less mug for handled mugs or a bottle with very different proportions, and measure pose accuracy in the coarse stage and the final output; if average 10-degree/5-cm accuracy does not fall substantially, the method is more tolerant to reference mismatch than the paper's premise, whereas a large drop would confirm the reference-sensitivity assumption. A second check: ablate the universal alignment loss $L_g$ on categories with large intra-category shape gaps (mugs and cameras) and compare 5-degree/2-cm accuracy; the loss is only causal if removing it degrades accuracy on those categories but not on shape-consistent ones.

Watch

Extended reading notes

Core claim

The central claim is that transferable semantic and geometric features alone—no fine-tuning, no instance-specific 3D models, no category-specific training—are sufficient for 6-DOF pose estimation on object categories the pipeline has never seen. The paper demonstrates this with a coarse-to-fine framework in which pre-trained 2D universal features (DINOv2 combined with Stable Diffusion) supply sparse correspondence candidates, Umeyama least-squares with RANSAC turns those into an initial pose, and iterative re-rendering repairs degraded correspondences when the pose gap is large. The fine stage uses pre-trained 3D universal features (DGCNN) in a newly proposed universal alignment loss that pulls high-confidence geometric correspondences together while mask and Chamfer losses keep the projection and shape plausible, allowing the reference mesh to deform and disambiguate pose from shape. On REAL275 and Wild6D, this method's best numbers are higher than the compared supervised, self-supervised, and zero-shot baselines on unseen categories, and ablations identify the iterative coarse stage and the 3D universal alignment loss as the main sources of improvement.

Load-bearing premise

The pipeline assumes that a reference mesh of the target category, selected in advance from Objaverse, is similar enough in shape and texture to the observed instance that 2D universal-feature correspondences can produce a coarse pose for the iterative loop to refine.

Editorial extensions

If this is right

A robot or AR system can estimate the pose of an object category it has never been trained on, given only an RGB-D frame and a generic reference mesh from an online repository.
The combination of DINOv2 and Stable Diffusion features yields better zero-shot correspondences than either feature alone, because SD supplies global context that DINOv2 lacks.
Iterating the reference render under the current pose recovers correspondences that would otherwise be lost when the initial pose is far off.
Adjusting the reference shape during refinement, anchored by 3D universal features, separates shape difference from pose error and improves final 6-DOF accuracy.
The pipeline also keeps high accuracy on the additional novel categories (teapot, tube) in the HouseCat6D test scenes, which the compared zero-shot methods largely fail on.

Reading between the lines

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

The confidence score used to choose among the four reference views could be repurposed as a model-selection criterion, letting a system pick among several candidate reference meshes automatically rather than assuming one good mesh per category.
Because the coarse stage already re-renders iteratively, the same render-and-compare loop could be extended past the refinement stage to add a second round of 2D correspondence repair after shape deformation, something the paper does not test.
The universal alignment loss is a generic geometric-semantic constraint; it could plausibly transfer to other category-level registration tasks in 3D, such as part alignment or canonicalization, though the paper only evaluates it for pose.
The method's 3.83 s per frame makes it a first-frame initialization strategy; the paper suggests tracking with fewer fitting steps, and a straightforward test would be to measure drift when the full pipeline runs only on the first frame and light refinement steps follow.

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.

Desk Editor's Note

read the letter

Quick take: this is a solid empirical paper that likely delivers the best zero-shot category-level pose numbers on REAL275 and Wild6D, but the word 'zero-shot' hides a manual reference-mesh selection step that the paper never specifies, and the benchmark numbers are single-run point estimates.

What is genuinely new: the two-stage pipeline that iteratively re-renders the reference model during coarse pose estimation to fix the correspondence degradation that plagues fixed-image DINO matching, then refines pose and shape together using 3D universal features from a pretrained DGCNN, with a universal alignment loss. Zero-Pose already showed DINO correspondences could do zero-shot category-level pose, and DINOv2/SD fusion was known for semantic matching, but the iterative re-rendering loop and the 3D geometric refinement with shape deformation are not in the prior work. The ablations in Table 2 do a good job of isolating each module's contribution.

What is soft: the reference-model selection. Section 3.1 just says a mesh is 'retrieved from Objaverse.' No retrieval rule, no scoring function, no failure analysis. The only robustness evidence is Table 4, which swaps in one alternative mesh per category; REAL275 IOU 0.25 drops from 80.06 to 75.08 and several Wild6D metrics shift by a few points. That tells me the method is not catastrophically sensitive, but the headline 'outperforms previous methods on unseen categories' is established for hand-picked references. A user with a genuinely new category has no guidance on what makes a good reference. Second, the hyperparameters (feature weights, loss weights, beta_g, fitting steps) are chosen on the same test benchmarks, and no error bars are provided. All reported numbers are single runs. Third, they reprocess Wild6D annotations and treat all mugs as symmetric; both are reasonable choices but are post-hoc evaluation decisions that should be flagged more prominently. No code is released, so the numbers are hard to verify independently.

Overall, this is a solid systems paper, not a breakthrough but a real step forward in zero-shot pose estimation. It deserves a serious referee and likely publication after a revision that specifies the reference-model selection protocol, adds error bars or a held-out hyperparameter selection, and releases code. I'd bring it to reading group.

Referee Report

3 major / 5 minor

Summary. The paper proposes a zero-shot category-level 6-DOF object pose estimation method from RGB-D input. It first uses pre-trained 2D universal features (DINOv2 and Stable Diffusion) to establish sparse correspondences between the target image and rendered images of a category-level reference mesh from Objaverse, producing a coarse pose through iterative Umeyama/RANSAC. It then refines pose and a per-vertex shape deformation by optimizing a loss that combines mask, Chamfer, and a new 3D universal-feature alignment term, using DGCNN features. The authors report state-of-the-art results on REAL275 and Wild6D under an unseen-category protocol, plus an additional HouseCat6D evaluation, and provide ablations for the iterative coarse stage, the refinement module, the feature combination, and the number of refinement steps.

Significance. If the results hold, the paper provides a strong zero-shot baseline for category-level pose estimation, demonstrating that transferable 2D and 3D features can replace category-specific training. The paper is commendable for its detailed ablations, the explicit evaluation on a novel dataset (HouseCat6D), and the inclusion of a reference-model robustness check in the appendix. The central claim, however, rests on two load-bearing assumptions that are not fully supported: a reproducible way to select the category reference mesh, and a validation protocol that does not tune hyperparameters on the test benchmarks. These issues do not invalidate the method but need to be addressed before the headline claim can be taken at face value.

major comments (3)

[Sec. 3.1, Sec. 3.2, Appendix Tab. 4] The reference mesh selection is not specified as a procedure. The text says only that the reference mesh is 'retrieved from Objaverse' (Sec. 3.1), with no retrieval algorithm, scoring function, or automatic selection criterion. This is load-bearing because the rendered reference images initialize all 2D-feature correspondences in Sec. 3.2, and a poor reference can produce a wrong coarse pose that the refinement stage is not designed to escape. The only robustness evidence is Appendix Tab. 4, which swaps in one alternative mesh per category: on REAL275 the IOU0.25 drops from 80.06 to 75.08, and on Wild6D from 88.46 to 85.38, with several pose metrics shifting by multiple points. The paper should provide a reproducible selection rule (e.g., an automatic retrieval criterion) or, failing that, report a distribution over multiple reference meshes and a failure-rate analysis. Without this, the zero-shot claim is only demonstrated for the authors' hand-picked references.
[Sec. 4.2, Tab. 1-6] The hyperparameters are selected on the same test benchmarks. Section 4.2 fixes alpha_D1, alpha_D2, alpha_SD, alpha_m, alpha_c, alpha_g, the number of coarse iterations, RANSAC iterations, and the refinement fitting steps, and Appendix Sec. 1.3 fixes alpha_p, alpha_ce, alpha_d; Table 6 then shows performance increasing with fitting steps up to the chosen 80. All benchmark numbers are single-run point estimates with no error bars or validation protocol. Since the central claim is quantitative superiority over prior methods, the reported margins (e.g., REAL275 10deg5cm: 57.74 vs 49.82 in Tab. 1) cannot be assessed for statistical or selection-bias significance. The authors should either adopt a validation-based hyperparameter selection protocol (e.g., a separate validation split per category) or report mean +/- std over multiple runs and show that the conclusions are robust to the choice of these hyperparameters.
[Appendix Sec. 2.1, Fig. 9] The treatment of mugs as symmetric objects is stated as follows: 'we treat mugs as symmetrical objects on all benchmarks, which will not affect the performance of comparison methods.' This is not self-evident: if symmetry-aware evaluation picks the closest pose among equivalent rotations, it can improve absolute scores for all methods, and the claim that it has no effect needs verification by reporting both symmetry-aware and symmetry-agnostic numbers for all baselines. At minimum, the paper should specify how symmetry equivalence is defined for mugs (e.g., which axis and angle tolerance) and confirm that the same evaluation code is used for every method.

minor comments (5)

[Sec. 3.2, Eq. (3)] The cyclical distance in Eq. (3) is written in a compressed notation that is hard to parse. Please define the nested argmax and the distance d(.,.) more explicitly, and clarify whether D is computed for all p or only for selected initial matches.
[Sec. 3.3] The line 'we abbreviate s_hat * R_hat as R_hat in Eq. 4' is confusing because Eq. (4) already solves for scale s. State the dimensions and the exact definition of the scale-corrected rotation used in the refinement optimization.
[Appendix Sec. 2.1] There is a typo: 'ans SSC-6D' should read 'and SSC-6D'. Also, the sentence 'For the category to be tested, we perform other category models on it' is awkward and should be rewritten.
[Sec. 4.4, Tab. 3] The row labeled 'ALL' on REAL275 reports higher IOU but lower pose accuracy than 'v2+SD'; the text explains this as boundary correspondences causing oversize scale. This interpretation would be more convincing with a qualitative example, as is already provided for other rows in Fig. 6.
[Sec. 4.2] The paper sets beta_g to 0.8 but does not ablate this threshold, despite the universal alignment loss being a claimed contribution. A small sensitivity study over beta_g would strengthen the argument.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: pose estimates come from frozen universal features, external reference meshes, and standard geometric solvers; benchmark results are independent empirical evidence.

full rationale

The paper's derivation chain is not circular. The coarse pose is computed by matching 2D features from rendered reference images and a target RGB-D image, lifting correspondences to 3D, and solving least-squares Umeyama with RANSAC (Eqs. 2-4); the refinement minimizes mask, Chamfer, and 3D feature alignment losses against the current pose. None of these equations is defined in terms of the pose it is claimed to predict, and no parameter is fitted to ground-truth pose labels; DINOv2, Stable Diffusion, and DGCNN are externally pretrained and used with frozen weights. The benchmark comparisons on REAL275, Wild6D, and HouseCat6D are external evaluations, and the paper includes an alternate-reference robustness check (Table 4). Mild caveats exist but are not circularity: some hyperparameters (feature weights, fitting steps, mug-symmetry convention) appear to be selected using the same test benchmarks, and the reference-mesh selection is not specified as an automatic procedure. These are methodological completeness and selection concerns, not equation-level reductions or self-citation chains. The paper also states its own limitations, such as runtime and occlusion sensitivity, which further indicates that the results are empirical rather than constructed.

Assumptions & free parameters 9 free parameters · 8 assumptions · 0 invented entities

The central claim rests on external pre-trained models, manually selected reference meshes, and several hand-set hyperparameters. No new physical or conceptual entities are introduced; the universal alignment loss is an objective function, not a postulated entity.

free parameters (9)

2D feature weights alpha_D1, alpha_D2, alpha_SD = 0, 0.7, 0.3
Set in Sec. 4.2 and selected by benchmark ablations (Tab. 3); they control which pre-trained feature contributes and directly affect correspondences and pose.
Pose refinement loss weights alpha_m, alpha_c, alpha_g = 1, 0.1, 1
Set in Sec. 4.2; chosen on the evaluation benchmarks with no held-out validation; they balance mask, Chamfer, and universal alignment terms.
Regularization weights alpha_p, alpha_ce, alpha_d = 20, 1, 1
Appendix Sec. 1.3; hand-set to keep the refined pose close to the initial pose and to constrain deformation, affecting the final pose-shape tradeoff.
Universal alignment threshold beta_g = 0.8
Sec. 3.3; filters low-confidence 3D feature pairs in the universal alignment loss; no sensitivity analysis is reported.
Number of coarse pose iterations = 2
Sec. 4.2 and Tab. 2 show that two iterations were chosen because they improve accuracy on the test benchmarks.
RANSAC iterations = 1000
Sec. 4.2; outlier handling hyperparameter with no sensitivity analysis reported.
Pose refinement fitting steps = 80
Appendix Tab. 6; chosen as a balance between time and accuracy from a sweep on the test benchmarks.
Keypoint correspondence count M = not reported
Eq. 3 and Eq. 4 use Top-k correspondences M, but the value is not given in the paper; this hidden choice affects coarse pose accuracy.
PCA feature dimension = 64
Appendix Sec. 1.2; reduces each universal feature before matching; chosen for computational convenience with no ablation shown.

assumptions (8)

domain assumption DINOv2 and Stable Diffusion features encode semantic correspondence that transfers across object categories and pose gaps.
Sec. 3.1 and Fig. 3; the method relies on this to establish coarse correspondences for unseen categories.
domain assumption DGCNN features pre-trained on point clouds provide geometric universal features that resolve pose-shape ambiguity for intra-category objects.
Sec. 3.3; used in the universal alignment loss Lg to constrain joint pose and shape optimization.
domain assumption Mask R-CNN provides reliable foreground masks and shape tokens for all test categories, including unseen ones.
Sec. 3.1 and Sec. 4.1; if segmentation fails, mask loss, feature extraction, and pose refinement degrade.
domain assumption A single reference mesh per category from Objaverse is an adequate shape prior for all instances of that category.
Sec. 3.2; the method renders four views of this mesh and uses it throughout; reference-model sensitivity is only lightly tested.
standard math Umeyama with RANSAC yields the correct rigid transform given noisy 2D-feature correspondences.
Eq. 4; standard robust least-squares approach for point-cloud alignment.
domain assumption Wild6D ground-truth poses can be reprocessed into NOCS coordinates without changing evaluation semantics.
Appendix Sec. 2.2; the authors manually unify annotations; errors in this step would affect all comparisons.
ad hoc to paper Mugs can be treated as symmetric objects for evaluation.
Appendix Sec. 2.1; chosen reference mug lacks a handle and Wild6D lacks handle annotations, so the symmetry convention changes how pose errors are scored.
domain assumption Regularization losses are sufficient to prevent degenerate shape deformation during refinement.
Sec. 3.3; there is no formal guarantee that the optimized deformation stays within plausible object shapes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Universal Features Guided Zero-Shot Category-Level Object Pose Estimation." pith.science (2026). https://pith.science/paper/XGBWQJJV

@misc{pith2026250102831,
  author       = {Pith},
  title        = {Pith review of: Universal Features Guided Zero-Shot Category-Level Object Pose Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XGBWQJJV}},
  note         = {Machine review of arXiv:2501.02831}
}

read the original abstract

Object pose estimation, crucial in computer vision and robotics applications, faces challenges with the diversity of unseen categories. We propose a zero-shot method to achieve category-level 6-DOF object pose estimation, which exploits both 2D and 3D universal features of input RGB-D image to establish semantic similarity-based correspondences and can be extended to unseen categories without additional model fine-tuning. Our method begins with combining efficient 2D universal features to find sparse correspondences between intra-category objects and gets initial coarse pose. To handle the correspondence degradation of 2D universal features if the pose deviates much from the target pose, we use an iterative strategy to optimize the pose. Subsequently, to resolve pose ambiguities due to shape differences between intra-category objects, the coarse pose is refined by optimizing with dense alignment constraint of 3D universal features. Our method outperforms previous methods on the REAL275 and Wild6D benchmarks for unseen categories.

Figures

Figures reproduced from arXiv: 2501.02831 by the authors.

**Figure 1.** (a) We propose a zero-shot pose estimation method for unseen categories using universal features and obtain accurate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Overview. Our framework includes a keypoint-level coarse pose estimation module and a pixel-level pose refinement [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Feature Performance Drop and Effect of Iterative [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: (a) Pose Refinement. Based on the coarse pose as [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Qualitative results on REAL275 and Wild6D. The red box represents the ground truth, and the green box represents [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 7.** Figure 7: (a) After the shape optimization, the reference ob [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 6.** Figure 6: Qualitative results on different combinations of 2D [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 8.** Figure 8: We render four reference images of each reference 3D model from the front, back, and two sides, as input to ’Coarse [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: Discussion of symmetric objects. The SD univer [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: We change the reference models from used in [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

We show the four views of the six category objects in Fig

Method Details 1.1 Initial View of Reference Images We render four reference images {Ir} of the reference 3D model from the front, back, and two sides. We show the four views of the six category objects in Fig. 8, respectively. They have great differences in texture and geometry from the in- stance objects in the benchmarks. 1.2 Details on Multi-Modal Uni...

work page 2023
[2]

Comparison Experiment Details 2.1 Baseline Implementation Details We compare our method with three types of baselines:

work page
[3]

2022) and MegaPose (Labb´e et al

Zero-Shot method: Zero-Pose (Goodwin et al. 2022) and MegaPose (Labb´e et al. 2023). We elaborate on the training and testing settings for the compared methods in Tab.1 of the main paper. We use the test sets of REAL275 (Wang et al. 2019b) and Wild6D (Ze and Wang 2022) for testing. For the supervised methods, we adapt the leave-1 strategy, which is to sel...

work page 2022
[5]

1.3 Details on Regularization Loss Lr

are not watertight, we use the method (Huang, Su, and Guibas 2018) to process these meshes into watertight meshes before feeding them into the pose refinement stage. 1.3 Details on Regularization Loss Lr. The regularization loss includes pose regularization loss Lp, center point regularization loss Lce, and deformation regu- larization loss Ld. We also fo...

work page 2018
[6]

Our regularization loss is defined as: Lr = αpLp + αceLce + αdLd

to minimize geometric distortion with normal, edge, and Laplacian constraints. Our regularization loss is defined as: Lr = αpLp + αceLce + αdLd. The pose regularization loss Lp enforces the refined pose to be close to the initial pose computed by the coarse esti- mation module, balancing the contributions between image and point cloud features: Lp = ∥( ¯R...

work page
[8]

2022a), VI- Net (Lin et al

Supervised methods: DPDN (Lin et al. 2022a), VI- Net (Lin et al. 2023) and SPD (Tian, Ang, and Lee 2020)

work page 2023
[9]

2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al

Self-Supervised methods: Self-Pose (Zhang et al. 2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al. 2022)

work page 2022
[11]

on synthetic CAMERA25 and real-world REAL275 datasets (Wang et al. 2019b). For the self-supervised methods Self-Pose (Zhang et al. 2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al. 2022), we directly use the officially trained model to con- duct the leave-p experiments. Because their solution requires prior information of category objects, and each...

work page 2022

Show all 14 references

[13]

2022), and the results are shown in Table 5

and Zero-Pose (Goodwin et al. 2022), and the results are shown in Table 5. For the comparison methods, we use the official pre- trained model for testing. We find that our method can still perform better when facing unseen categories because our method achieves pose estimation...

2022
[14]

10, and we show the re- sults in Tab

Additional Ablation Experiments 3.1 Effect of Reference 3D Model To demonstrate that our pipeline is not sensitive to the spe- cific geometry and texture of the reference 3D model, we replace the objects used in the main paper with significantly different objects as shown in F...

1981
[2017]

arXiv preprint arXiv:1711.00199

Posecnn: A convolutional neural network for 6d ob- ject pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199. Ze, Y .; and Wang, X. 2022. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. Advances in Ne...

2022 arXiv
[2020]

For 2D uni- versal feature extraction, we follow the methods (Goodwin et al

to render the reference RGB-D images. For 2D uni- versal feature extraction, we follow the methods (Goodwin et al. 2022; Zhang et al. 2023; Luo et al. 2023) and use DI- NOv1(‘vit small’), DINOv2(‘vits14’) and SD(‘v1-5’). The 3D universal features are extracted from conv6 of th...

2022
[2022]

2019b), which is quite different from the annotations of REAL275 (Wang et al

test set are not aligned to the NOCS coordinate (Wang et al. 2019b), which is quite different from the annotations of REAL275 (Wang et al. 2019b). So we perform low-cost processing on it and unify the annotations into the NOCS coordinate. 2.3 Additional Comparison Experiment T...

2024
[2023]

arXiv preprint arXiv:2305.17934, 2

ZeroPose: CAD-model-based zero-shot pose estima- tion. arXiv preprint arXiv:2305.17934, 2. Chen, W.; Jia, X.; Chang, H. J.; Duan, J.; Shen, L.; and Leonardis, A. 2021. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled ro- tation mecha...

2021 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

We show the four views of the six category objects in Fig

Method Details 1.1 Initial View of Reference Images We render four reference images {Ir} of the reference 3D model from the front, back, and two sides. We show the four views of the six category objects in Fig. 8, respectively. They have great differences in texture and geometry from the in- stance objects in the benchmarks. 1.2 Details on Multi-Modal Uni...

work page 2023

[2] [2]

Comparison Experiment Details 2.1 Baseline Implementation Details We compare our method with three types of baselines:

work page

[3] [3]

2022) and MegaPose (Labb´e et al

Zero-Shot method: Zero-Pose (Goodwin et al. 2022) and MegaPose (Labb´e et al. 2023). We elaborate on the training and testing settings for the compared methods in Tab.1 of the main paper. We use the test sets of REAL275 (Wang et al. 2019b) and Wild6D (Ze and Wang 2022) for testing. For the supervised methods, we adapt the leave-1 strategy, which is to sel...

work page 2022

[4] [5]

1.3 Details on Regularization Loss Lr

are not watertight, we use the method (Huang, Su, and Guibas 2018) to process these meshes into watertight meshes before feeding them into the pose refinement stage. 1.3 Details on Regularization Loss Lr. The regularization loss includes pose regularization loss Lp, center point regularization loss Lce, and deformation regu- larization loss Ld. We also fo...

work page 2018

[5] [6]

Our regularization loss is defined as: Lr = αpLp + αceLce + αdLd

to minimize geometric distortion with normal, edge, and Laplacian constraints. Our regularization loss is defined as: Lr = αpLp + αceLce + αdLd. The pose regularization loss Lp enforces the refined pose to be close to the initial pose computed by the coarse esti- mation module, balancing the contributions between image and point cloud features: Lp = ∥( ¯R...

work page

[6] [8]

2022a), VI- Net (Lin et al

Supervised methods: DPDN (Lin et al. 2022a), VI- Net (Lin et al. 2023) and SPD (Tian, Ang, and Lee 2020)

work page 2023

[7] [9]

2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al

Self-Supervised methods: Self-Pose (Zhang et al. 2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al. 2022)

work page 2022

[8] [11]

on synthetic CAMERA25 and real-world REAL275 datasets (Wang et al. 2019b). For the self-supervised methods Self-Pose (Zhang et al. 2022), Wild6D (Ze and Wang 2022) ans SSC-6D (Peng et al. 2022), we directly use the officially trained model to con- duct the leave-p experiments. Because their solution requires prior information of category objects, and each...

work page 2022