Pith. sign in

REVIEW 3 major objections 5 minor 47 references

One model can recover an object's full 9D pose from a single masked view without category labels, CAD models, or reference images.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 01:14 UTC pith:WXESO26S

load-bearing objection Solid category-agnostic 9D pose system with real multi-benchmark transfer; the “beats specialists” line is partly protocol-driven and should stay qualified. the 3 major comments →

arxiv 2607.09985 v1 pith:WXESO26S submitted 2026-07-10 cs.CV

UniPose9D: Universal Category-Agnostic Object Pose Estimation

Yang You , Yi Du , Cole Harrison , Leonidas Guibas This is my paper
classification cs.CV
keywords object pose estimationcategory-agnostic9D poseNOCSpoint-pair featuresKabsch-Umeyamaflow matchingRGB-D
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most pose estimators either memorize specific objects with CAD models or stay locked inside a fixed category taxonomy with mean-shape priors. UniPose9D argues that neither is necessary: a single observation of a masked object, plus measured or predicted depth, already carries enough appearance and geometry cues to recover rotation, translation, and metric box size. The system samples point pairs from the visible surface, predicts their positions in a shared canonical cube (NOCS) with visual and geometric features, then aligns those predictions to the observed cloud with a scale-aware, multi-hop RANSAC solver. Flow matching handles symmetries that would otherwise produce multiple valid poses. Trained once on a curated mix of public datasets, the same weights match or beat specialist models on standard benchmarks and transfer to entirely unseen objects and in-the-wild photographs.

Core claim

A single category-agnostic network, given only an instance mask/ROI and RGB-D (or RGB with predicted metric depth), can estimate full 9D pose—rotation, translation, and metric size—without category labels, CAD models, mean-shape priors, or reference views, and can match or surpass specialist methods while generalizing to unseen objects and scenes.

What carries the argument

Point-pair NOCS prediction plus pairwise metric-scale calibration and an adaptive N-hop Kabsch–Umeyama RANSAC: pairs give quadratic correspondences and a direct ratio for scale; the multi-hop solver then tightens the inlier threshold from the running size estimate so the final SE(3) alignment is recovered from the observed geometry itself.

Load-bearing premise

Most objects have a human-aligned canonical orientation that a single partial view can still reveal from appearance and geometry alone, without a category template or mean shape.

What would settle it

On a held-out set of everyday objects whose natural upright orientation is ambiguous or contested (or under severe occlusion and thin structure), check whether UniPose9D's predicted axes and metric boxes remain consistent with human canonical labels and with ground-truth pose; systematic failure would collapse the claim that one observation suffices.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. UniPose9D proposes a single category-agnostic model for 9D object pose estimation (rotation, translation, and metric size) from a masked RGB-D observation or RGB with predicted depth, without category labels, CAD models, mean-shape priors, or reference views at inference. The pipeline samples point pairs from the observed object cloud, fuses DINOv2 visual features with multi-radius PointNet geometry, predicts pairwise NOCS coordinates via a Gaussian-mixture flow-matching head plus a metric box-size head, then recovers pose with pairwise scale calibration and an adaptive N-hop Kabsch–Umeyama RANSAC. Training mixes curated pose annotations from PACE, Omni6DPose, NOCS REAL275, and HouseCat6D. Empirically, one shared model is best or near-best on PACE, Omni6DPose, and HouseCat6D, competitive on REAL275 against specialists, and stronger than GenPose++ / Orient Anything on unseen YCB-Video and HOPE under RGB and RGB-D.

Significance. If the results hold under fair comparison, this is a useful step toward a reusable 9D pose foundation model for robotics and AR: full metric pose from a single masked observation without per-category priors or reference assets. Strengths include multi-benchmark evaluation (Tables 1–5), explicit Gen vs CS baseline tagging, ablations of PointNet, pairs, flow matching, scale decoupling, and N-hop (Table 6), a unified-training diagnostic (Supp. Table S2), failure cases (Fig. S6), and public code/model. The technical combination of pairwise NOCS, pairwise scale calibration (Eq. 10), adaptive N-hop Umeyama, and flow matching for symmetry is concrete and falsifiable. The main significance risk is attribution: how much of the “match/surpass specialists” headline is architecture versus multi-dataset curation and protocol differences.

major comments (3)
  1. Abstract / §1 / Tables 1–4 vs Supp. Table S2: the central claim that one category-agnostic model can match or surpass specialists is only partly controlled. In-domain tables largely compare UniPose9D (four-dataset mixture) to CS methods under native single-taxonomy protocols. Supp. Table S2 is the right diagnostic, but incomplete: under All-DS, SecondPose’s REAL275 5°2cm falls from 56.2 to 36.2 while UniPose9D reports 45.4, and GenPose++ improves on some transfers yet still lags. This is consistent with specialists being hurt by forced multi-dataset training and/or UniPose9D gains coming largely from curated alignment and balanced sampling (§3.6, Supp. C). Please expand the apples-to-apples protocol (same mixture, same mask/depth inputs, no per-category branches where possible) and report native vs All-DS for the main metrics, or qualify the headline as “competitive under a shared multi-
  2. §1 and §3.1 (NOCS / canonical orientation premise): the method assumes a human-aligned, unique or finite-set canonical frame recoverable from a single partial observation without category templates. After “curating and aligning pose annotations” across heterogeneous datasets, the paper should document how canonical axes/sizes were made consistent (especially for symmetric or near-symmetric objects), and report failure rates or multi-mode statistics when the frame is ambiguous. Without that, pairwise NOCS + RANSAC can still output a pose that is not uniquely meaningful; Fig. S6 already shows occlusion/thin-structure failures, but the curation protocol is load-bearing for the category-agnostic claim.
  3. §4.1 / Table 5: RGB results depend on MoGe-2 predicted metric depth, and translation Acc@2cm is weak (3.4 / 1.5 on YCB / HOPE). The paper correctly notes depth uncertainty, but the 9D claim for RGB-only is only partially supported. Please separate rotation vs translation/size more clearly under predicted depth, and state whether metric size is recovered from predicted depth geometry or mainly from the learned box head; otherwise the RGB setting risks overstating full 9D reliability.
minor comments (5)
  1. §3.5 / Algorithm 1: specify the initial threshold τ0, RANSAC iteration budget, and how pair-level inliers are mapped to the flattened correspondence list used for final Umeyama.
  2. §3.4 / Eq. (5): clarify the exact definition of “norm(·)” (which scale statistic of the ROI cloud) and whether it is recomputed after DZI augmentation.
  3. §3.6 and Supp. B: report approximate training wall-clock, number of point pairs T per image, and inference latency so the foundation-model claim is operationally grounded.
  4. Figure 3 / qualitative figures: add a short note on whether boxes are drawn from predicted s or from GT size when comparing rotation-only baselines (as done for Orient Anything in the supplement).
  5. Related work: CPPF/CPPF++ are acknowledged; a one-sentence contrast on category-specific training vs the shared multi-dataset protocol would help readers place the contribution.

Circularity Check

0 steps flagged

No significant circularity: UniPose9D is a supervised learning pipeline evaluated on held-out and external benchmarks, not a derivation that forces its 9D scores by construction.

full rationale

The paper’s load-bearing claims are empirical: a single model trained on curated public pose annotations predicts pairwise NOCS and metric box size, then recovers (R, t, s) via scale-calibrated N-hop Kabsch–Umeyama, and is scored under standard IoU/AP/VUS protocols on held-out PACE/Omni6DPose/REAL275/HouseCat6D splits plus unseen YCB-Video and HOPE. Equations (1), (9)–(12) are geometric alignment identities relating predicted NOCS to observed points; they do not define the reported metrics in terms of fitted targets. Author-overlapping priors (CPPF/CPPF++ for point-pair voting, PACE as a training source, 3DCorrEnhance for DINOv2 features, Orient Anything for the informal canonical-orientation premise) supply methods, data, or motivation but are not uniqueness theorems and do not algebraically force the Table 1–5 numbers. Unified multi-dataset training may confound “architecture vs. data mixture” comparisons (Supp. Table S2), but that is an experimental-design concern, not circular reduction of prediction to input. No self-definitional step, fitted-parameter-as-prediction, or load-bearing self-citation chain is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard 3D-vision modeling choices (NOCS, SE(3) alignment, RANSAC), the domain premise that objects have recoverable human-aligned canonical frames from one view, availability of masks and metric depth, and many training/inference hyperparameters. No new physical entity is postulated; the main invented constructs are algorithmic (N-hop adaptive solver, pair-level multimodal NOCS head in this protocol).

free parameters (6)
  • N-hop count H = 3
    Number of adaptive RANSAC refinement hops; set to H=3 in §3.5 without a uniqueness derivation.
  • Adaptive inlier threshold schedule = 0.05 scale factor; 10% support; 1.5 relax
    τ_{h+1}=0.05·mean_inlier_∞-box-size if ≥10% inliers, else 1.5·τ_h (§3.5 / Alg. 1); hand-chosen constants that control pose recovery.
  • PointNet grouping radii = {0.02, 0.04, 0.08}
    Multi-radius neighborhoods {0.02,0.04,0.08} chosen for geometric features (§3.6); sensor/scale dependent.
  • Loss weight λ_scale = 1
    Balances NOCS flow NLL and box-size L2; set to 1 (§3.6).
  • Gaussian mixture components G and ODE sampler steps = G=32; 16 steps × 2 substeps
    Flow head uses 32-component mixture; inference uses 16 ODE steps with 2 substeps (Supplement B / §3.6)—architecture/inference knobs affecting multimodal pose recovery.
  • AdamW learning rate and training schedule = 1e-4; 30 epochs; batch 16
    Base LR 1e-4, cosine decay, 30 epochs, batch 16, balanced multi-dataset sampling—standard fitted training hyperparameters (§3.6, Supp. B.2).
axioms (5)
  • domain assumption Most objects have a naturally defined human-aligned canonical orientation (unique or finite under symmetry) recoverable from a single observation.
    Stated in §1 via Orient Anything; underpins category-agnostic NOCS prediction without templates.
  • domain assumption NOCS maps an object into a unit cube with longest side 1 and category-consistent orientation; metric pose follows p=γ R x + t.
    §3.1 Preliminary; standard NOCS modeling assumption used for supervision and recovery.
  • domain assumption An instance mask/ROI is available at inference, and depth is metric (sensor depth or off-the-shelf predicted metric depth).
    §4.1 setup and §5 limitations; without this the point cloud and scale are undefined.
  • standard math Kabsch–Umeyama / RANSAC recovers SE(3) from noisy correspondences when enough inliers exist.
    §3.5 pose recovery; classical rigid registration assumption.
  • ad hoc to paper Pose annotations from heterogeneous public datasets can be curated/aligned into one consistent training mixture for a shared model.
    Abstract and §1/§3.6; quality of cross-dataset canonical alignment is paper-specific and not independently audited in the text.
invented entities (2)
  • N-hop Kabsch–Umeyama with adaptive inlier threshold no independent evidence
    purpose: Iteratively refine rigid alignment and scale-aware thresholds from point-pair NOCS correspondences.
    Algorithmic construct introduced in §3.5 / Alg. 1; evaluated only inside this pipeline, not as an independent theorem with external certificates.
  • Point-pair multimodal NOCS + metric box head (UniPose9D pipeline) no independent evidence
    purpose: Predict pairwise canonical coordinates and sizes from DINOv2+PointNet features under category-agnostic inference.
    System-level invention combining known modules; evidence is empirical benchmark performance, not an external physical prediction.

pith-pipeline@v1.1.0-grok45 · 20783 in / 4162 out tokens · 40971 ms · 2026-07-14T01:14:58.196760+00:00 · methodology

0 comments
read the original abstract

Object pose estimation is a fundamental problem in 3D vision. Although recent state-of-the-art approaches achieve strong performance, they often overfit to existing benchmarks and exhibit limited generalization to novel categories and unseen scenes. We propose UniPose9D, a category-agnostic foundation model for 9D object pose estimation: given an instance mask/ROI and either an RGB-D observation or an RGB image with predicted depth, the model estimates rotation, translation, and metric size without category labels, CAD models, mean-shape priors, or reference views. Specifically, UniPose9D samples point pairs from the observed object geometry and uses DINOv2 and PointNet features to predict NOCS coordinates for each pair. To improve accuracy, we introduce a point-pair-based RANSAC N-hop Kabsch--Umeyama algorithm with an adaptive threshold. We further employ flow matching to address symmetric ambiguities and construct a large-scale training set by curating and aligning pose annotations from existing public datasets. Experiments across six datasets show that a single unified model can match or surpass specialist methods while generalizing to unseen objects and in-the-wild scenarios. Our code and model are available on https://github.com/qq456cvb/UniPose9D.

Figures

Figures reproduced from arXiv: 2607.09985 by Cole Harrison, Leonidas Guibas, Yang You, Yi Du.

Figure 1
Figure 1. Figure 1: UniPose9D. Given an instance mask/ROI and either RGB-D input (left four examples) or RGB input with predicted depth (right three examples), we predict a full 9D pose and metric 3D bounding box without category labels, CAD models, mean-shape priors, or reference views. Examples include cluttered tabletops, severe occlusions, a robotic manipulation scene, and in-the-wild photographs of novel objects. Yellow … view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of UniPose9D. (A) Feature extraction: given an RGB crop I and point cloud P, DINOv2 and a lightweight PointNet produce per-point descriptors. For each sampled point pair T =(i1, i2), we build fT by concatenating descriptors and relative coordinates. (B) Point-pair sampling and NOCS prediction: an MLP with flow matching predicts pairwise NOCS coordinates xˆi1 , xˆi2 and a per-pair metric box size ˆ… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on PACE. Against NOCS, HS-Pose, CPPF++, and GenPose++, our method yields tighter 3D boxes and more consistent axes. PACE [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 8 linked inside Pith

  1. [1]

    Gaussian mixture flow matching models

    Hansheng Chen, Kai Zhang, Hao Tan, Zexiang Xu, Fujun Luan, Leonidas Guibas, Gordon Wetzstein, and Sai Bi. Gaussian mixture flow matching models. InICML, 2025

    2025

  2. [2]

    Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation

    Kai Chen and Qi Dou. Sgpa: Structure-guided prior adaptation for category-level 6d object pose estimation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2773–2782, 2021

    2021

  3. [3]

    Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism

    Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Linlin Shen, and Ales Leonardis. Fs-net: Fast shape-based network for category-level 6d object pose estimation with decoupled rotation mechanism. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1581–1590, 2021

    2021

  4. [4]

    Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation

    Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959–9969, 2024

    2024

  5. [5]

    Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting

    Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, and Federico Tombari. Gpv-pose: Category-level object pose estimation via geometry-guided point-wise voting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022

    2022

  6. [6]

    Model globally, match locally: Efficient and robust 3d object recognition

    Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In2010 IEEE computer society conference on computer vision and pattern recognition, pages 998–1005. Ieee, 2010

    2010

  7. [7]

    Guoguang Du, Kai Wang, Shiguo Lian, and Kaiyong Zhao. Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review.Artificial Intelligence Review, 54(3):1677–1734, 2021

    2021

  8. [8]

    Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings

    Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6749–6758, 2022

    2022

  9. [9]

    Onepose++: Keypoint- free one-shot object pose estimation without cad models.Advances in Neural Information Processing Systems, 35:35103–35115, 2022

    Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint- free one-shot object pose estimation without cad models.Advances in Neural Information Processing Systems, 35:35103–35115, 2022

    2022

  10. [10]

    Fs6d: Few-shot 6d pose estimation of novel objects

    Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6814–6824, 2022. 10

    2022

  11. [11]

    Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios

    HyunJun Jung, Shun-Cheng Wu, Patrick Ruhkamp, Guangyao Zhai, Hannah Schieber, Giulia Rizzoli, Pengyuan Wang, Hongcheng Zhao, Lorenzo Garattoni, Sven Meier, et al. Housecat6d-a large-scale multi-modal category level 6d object perception dataset with household objects in realistic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and P...

    2024

  12. [12]

    Cosypose: Consistent multi-view multi- object 6d pose estimation

    Yann Labbé, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi- object 6d pose estimation. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16, pages 574–591. Springer, 2020

    2020

  13. [13]

    A purely algebraic justification of the kabsch- umeyama algorithm.Journal of Research of the National Institute of Standards and Technology, 124: 124028, 2019

    Jim Lawrence, Javier Bernal, and Christoph Witzgall. A purely algebraic justification of the kabsch- umeyama algorithm.Journal of Research of the National Institute of Standards and Technology, 124: 124028, 2019. doi: 10.6028/jres.124.028. URLhttps://doi.org/10.6028/jres.124.028

    doi:10.6028/jres.124.028 2019

  14. [14]

    Any6D: Model-free 6d pose estimation of novel objects

    Taeyeop Lee, Bowen Wen, Minjun Kang, Gyuree Kang, In So Kweon, and Kuk-Jin Yoon. Any6D: Model-free 6d pose estimation of novel objects. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025

    2025

  15. [15]

    Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation

    Haitao Lin, Zichang Liu, Chilam Cheang, Yanwei Fu, Guodong Guo, and Xiangyang Xue. Sar-net: Shape alignment and recovery network for category-level 6d object pose and size estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717, 2022

    2022

  16. [16]

    Jiehong Lin, Hongyang Li, Ke Chen, Jiangbo Lu, and Kui Jia. Sparse steerable convolutions: An efficient learning of se (3)-equivariant features for estimation and tracking of object poses in 3d space.Advances in Neural Information Processing Systems, 34:16779–16790, 2021

    2021

  17. [17]

    Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency

    Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. Dualposenet: Category-level 6d object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3560–3569, 2021

    2021

  18. [18]

    Vi-net: Boosting category-level 6d object pose estimation via learning decoupled rotations on the spherical representations

    Jiehong Lin, Zewei Wei, Yabin Zhang, and Kui Jia. Vi-net: Boosting category-level 6d object pose estimation via learning decoupled rotations on the spherical representations. InProceedings of the IEEE/CVF international conference on computer vision, pages 14001–14011, 2023

    2023

  19. [19]

    Instance-adaptive and geometric-aware keypoint learning for category-level 6d object pose estimation

    Xiao Lin, Wenfei Yang, Yuan Gao, and Tianzhu Zhang. Instance-adaptive and geometric-aware keypoint learning for category-level 6d object pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21040–21049, 2024

    2024

  20. [20]

    Ist-net: Prior-free category-level pose estimation with implicit space transformation

    Jianhui Liu, Yukang Chen, Xiaoqing Ye, and Xiaojuan Qi. Ist-net: Prior-free category-level pose estimation with implicit space transformation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13978–13988, 2023

    2023

  21. [21]

    Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images

    Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. InEuropean Conference on Computer Vision, pages 298–315. Springer, 2022

    2022

  22. [22]

    Pose estimation for augmented reality: a hands-on survey.IEEE transactions on visualization and computer graphics, 22(12):2633–2651, 2015

    Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey.IEEE transactions on visualization and computer graphics, 22(12):2633–2651, 2015

    2015

  23. [23]

    Project tango

    Eitan Marder-Eppstein. Project tango. InACM SIGGRAPH 2016 Real-Time Live!, pages 25–25. 2016

    2016

  24. [24]

    Nope: Novel object pose estimation from a single image

    Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. Nope: Novel object pose estimation from a single image. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17923–17932, 2024

    2024

  25. [25]

    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

    2024

  26. [26]

    Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Laba...

    2023

  27. [27]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2017. 11

    2017

  28. [28]

    Maskfusion: Real-time recognition, tracking and reconstruction of multiple moving objects

    Martin Runz, Maud Buffier, and Lourdes Agapito. Maskfusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In2018 IEEE international symposium on mixed and augmented reality (ISMAR), pages 10–20. IEEE, 2018

    2018

  29. [29]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021

    2021

  30. [30]

    Onepose: One-shot object pose estimation without cad models

    Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825–6834, 2022

    2022

  31. [31]

    6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark

    Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-dof pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. InInternational Conference on Intelligent Robots and Systems (IROS), 2022

    2022

  32. [32]

    GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation

    Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16611–16621, June 2021

    2021

  33. [33]

    Normal- ized 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. Normal- ized 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 2642–2651, 2019

    2019

  34. [34]

    Moge-2: Accurate monocular geometry with metric scale and sharp details

    Ruicheng Wang, Sicheng Xu, Yue Dong, Yu Deng, Jianfeng Xiang, Zelong Lv, Guangzhong Sun, Xin Tong, and Jiaolong Yang. Moge-2: Accurate monocular geometry with metric scale and sharp details. arXiv preprint arXiv:2507.02546, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Orient anything: Learning robust object orientation estimation from rendering 3d models.arXiv preprint arXiv:2412.18605, 2024

    Zehan Wang, Ziang Zhang, Tianyu Pang, Chao Du, Hengshuang Zhao, and Zhou Zhao. Orient anything: Learning robust object orientation estimation from rendering 3d models.arXiv preprint arXiv:2412.18605, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    You only demonstrate once: Category-level manipulation from single visual demonstration.arXiv preprint arXiv:2201.12716, 2022

    Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration.arXiv preprint arXiv:2201.12716, 2022

    Pith/arXiv arXiv 2022

  37. [37]

    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

    2024

  38. [38]

    Segicp: Integrated deep semantic segmentation and pose estimation

    Jay M Wong, Vincent Kee, Tiffany Le, Syler Wagner, Gian-Luca Mariottini, Abraham Schneider, Lei Hamilton, Rahul Chipalkatty, Mitchell Hebert, David MS Johnson, et al. Segicp: Integrated deep semantic segmentation and pose estimation. In2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5784–5789. IEEE, 2017

    2017

  39. [39]

    Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes.arXiv preprint arXiv:1711.00199, 2017

    Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes.arXiv preprint arXiv:1711.00199, 2017

    Pith/arXiv arXiv 2017

  40. [40]

    Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation.arXiv preprint arXiv:2211.13398, 2022

    Yang You, Wenhao He, Jin Liu, Hongkai Xiong, Weiming Wang, and Cewu Lu. Cppf++: Uncertainty-aware sim2real object pose estimation by vote aggregation.arXiv preprint arXiv:2211.13398, 2022

    Pith/arXiv arXiv 2022

  41. [41]

    Cppf: Towards robust category-level 9d pose estimation in the wild

    Yang You, Ruoxi Shi, Weiming Wang, and Cewu Lu. Cppf: Towards robust category-level 9d pose estimation in the wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6866–6875, 2022

    2022

  42. [42]

    Multiview equivariance improves 3d correspondence understanding with minimal feature finetuning.arXiv preprint arXiv:2411.19458, 2024

    Yang You, Yixin Li, Congyue Deng, Yue Wang, and Leonidas Guibas. Multiview equivariance improves 3d correspondence understanding with minimal feature finetuning.arXiv preprint arXiv:2411.19458, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Pace: A large-scale dataset with pose annotations in cluttered environments

    Yang You, Kai Xiong, Zhening Yang, Zhengxiang Huang, Junwei Zhou, Ruoxi Shi, Zhou Fang, Adam W Harley, Leonidas Guibas, and Cewu Lu. Pace: A large-scale dataset with pose annotations in cluttered environments. InEuropean Conference on Computer Vision, pages 473–489. Springer, 2024

    2024

  44. [44]

    Multi-view self-supervised deep learning for 6d pose estimation in the amazon picking challenge

    Andy Zeng, Kuan-Ting Yu, Shuran Song, Daniel Suo, Ed Walker, Alberto Rodriguez, and Jianxiong Xiao. Multi-view self-supervised deep learning for 6d pose estimation in the amazon picking challenge. In2017 IEEE international conference on robotics and automation (ICRA), pages 1386–1383. IEEE, 2017

    2017

  45. [45]

    Genpose: Generative category-level object pose estimation via diffusion models.arXiv preprint arXiv:2306.10531, 2023

    Jiyao Zhang, Mingdong Wu, and Hao Dong. Genpose: Generative category-level object pose estimation via diffusion models.arXiv preprint arXiv:2306.10531, 2023. 12

    Pith/arXiv arXiv 2023

  46. [46]

    Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking.arXiv preprint arXiv:2406.04316, 2024

    Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking.arXiv preprint arXiv:2406.04316, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Hs-pose: Hybrid scope feature extraction for category-level object pose estimation

    Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Aleš Leonardis, Wei Zhang, and Hyung Jin Chang. Hs-pose: Hybrid scope feature extraction for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17163– 17173, 2023. 13 Supplementary Material Figure S1:Qualitative c...

    2023