Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Box Pose and Shape Estimation and Domain Adaptation for Large-Scale Warehouse Automation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims a self-supervised 'correct-and-certify' pipeline lets warehouse robots estimate box pose and shape from unlabeled stereo images, cutting real-world position error from 2.080 m to 0.134 m.

desk verdict A solid self-supervised sim-to-real pipeline with independent keypoint evidence; the circular pseudo-GT evaluation and unmeasured SAM2-certificate precision are the main caveats. read the letter →

arxiv 2507.00984 v1 pith:HHSDABRQ submitted 2025-07-01 cs.RO cs.CVcs.LG

Xihang Yu , Rajat Talak , Jingnan Shi , Ulrich Viereck , Igor Gilitschenski , Luca Carlone This is my paper
classification cs.ROcs.CVcs.LG
keywords Self-supervisedlearningObjectposeestimationShapeDomainadaptationWarehouseautomationCertifiablemodelsStereovisionKeypointdetection
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 tries to establish that a robot's box pose-and-shape perception can be adapted from simulation to real warehouse conditions using only unlabeled images collected by the robots themselves. The proposed pipeline, named BOSS, first predicts box-corner keypoints from stereo images, then solves a two-view perspective-n-point problem to recover pose and shape, and finally self-trains only on keypoints that pass three certificates (mask overlap, reprojection residual, epipolar consistency). On a real dataset of 9,000 images, the self-supervised model reduces average position error from 2.080 m (simulation-only) to 0.134 m, approaching the supervised upper bound of 0.111 m; a larger unlabeled set of 50,000 images gives a small further gain. If true, warehouse fleets can continuously improve their perception from raw operational data, without the cost of manual annotation.

What carries the argument

The machinery is the correct-and-certify self-training loop. A stereo pair of images feeds two keypoint networks; only high-confidence corners are kept and fed into a two-view perspective-n-point optimization that solves for pose $T \in SE(3)$ and shape $S = \mathrm{diag}(a,b,c)$ while down-weighting outliers with a Geman-McClure robust loss. The estimated model is reprojected and admitted as a pseudo-label only if it passes three certificates: the 2D mask IoU certificate of Eq. (2), the residual certificate of Eq. (3), and the epipolar constraint certificate of Eq. (4). When ground-truth masks are absent, a segmentation model yields pseudo-masks using uniform simplex sampling of polygon interiors, chosen because it covers the projected box polygon uniformly.

What would settle it

Take a set of real warehouse stereo images with ground-truth keypoints and run the pipeline while artificially degrading the segmentation masks, for example by eroding or dilating them. If the mask-IoU certificate continues to admit pseudo-labels whose keypoint RMSE exceeds the acceptable threshold, or if the self-trained model's final position error grows materially relative to the ground-truth-mask variant, that would falsify the assumption that the segmentation model is a reliable substitute for ground truth.

Watch

Extended reading notes

Core claim

The central discovery is that the correct-and-certify paradigm transfers to industrial box perception: corrected pseudo-labels, filtered by certifiable checks, are good enough to close most of the sim-to-real gap. The paper shows that self-training on unlabeled real stereo images, using a keypoint predictor pretrained only in simulation, yields an average position error of 0.134 m and rotation error of 0.223 rad on a 9,000-image real dataset, compared with 2.080 m and 0.554 rad for the simulation-only model, while the fully supervised upper bound is 0.111 m and 0.212 rad. When the 2D certificate uses masks from a segmentation model rather than ground-truth masks, performance remains nearly identical, and filtering by the certificates at inference brings the model close to the supervised baseline.

Load-bearing premise

The load-bearing premise is that the segmentation masks produced on real warehouse images are accurate enough that the mask-overlap certificate keeps only good pseudo-labels, even though the certificate was validated with ground-truth masks and the segmentation model's error on warehouse images is not directly measured.

Editorial extensions

If this is right

  • Warehouse robots can retrain their perception models on their own operational footage, with no human annotation, so perception improves as the fleet runs.
  • Models that start from simulation and adapt with this pipeline reach close to the accuracy of models trained on fully labeled real data.
  • Adding 50,000 unlabeled images yields a small but consistent accuracy gain over 9,000, suggesting that self-training benefits from more data.
  • Using learned segmentation masks instead of ground-truth masks in the certificate costs little accuracy, removing the main annotation requirement.
  • Certificate filtering applied at inference time, not only during training, further improves keypoint accuracy and approaches the supervised baseline.

Reading between the lines

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

  • The authors leave implicit that an automatic, adaptive threshold schedule could replace the fixed certificate thresholds, and the reported limited gain from scaling to 50k images hints that threshold rigidity, not data scarcity, is the current bottleneck.
  • A natural testable extension is to apply the same correct-and-certify loop to other cuboid-like objects (pallets, totes, crates) where stereo keypoints are easy to define and masks come from the same segmentation model.
  • Because the pipeline's reliability hinges on segmentation-mask quality, one could build a monitoring statistic over the certificate pass-rate in deployment to detect when the mask model drifts, although the paper does not discuss such a safeguard.
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

4 major / 6 minor

Summary. The paper presents BOSS, a self-supervised pipeline for estimating the pose and shape of boxes from stereo RGB images in warehouse environments. A stereo Keypoint-RCNN detects box-corner keypoints; a two-view Perspective-n-Point estimator recovers pose and cuboid dimensions; and a correct-and-certify self-training loop uses three certificates (2D IoU, residual, and epipolar) to select pseudo-labels from unlabeled data. The method is evaluated in sim2sim experiments, on the annotated Symbotic-9k real dataset, and on a 50k-image unlabeled dataset. Results show that the self-supervised models substantially outperform simulation-only training and approach a supervised upper bound on keypoint metrics, while also beating a zero-shot 3D bounding-box baseline (Cube R-CNN).

Significance. If the results hold, this is a strong industrial demonstration of self-supervised domain adaptation for object pose/shape estimation, with clear practical value for warehouse automation. The paper's strengths include: evaluation on a real annotated dataset (Symbotic-9k) with keypoint-level metrics; a large-scale 50k-image self-training study; comparison against a strong zero-shot baseline; and a systematic empirical analysis of certificates and SAM2 sampling strategies. The central claim that self-supervised adaptation improves over sim-only training is independently supported by the annotated keypoint test set. However, the certificate validation gap for SAM2 masks and the use of pseudo ground truth for real pose/shape metrics are load-bearing limitations that require attention before the claims are fully established.

major comments (4)
  1. [Section 4.1 and 4.4] The 2D certificate in Eq. (2) is validated only with ground-truth masks (Fig. 4a), and the SAM2 sampling study (Fig. 3) is conducted on a synthetic 2D octagon rather than on warehouse images. In the actual self-training loop, the mask M in Eq. (2) is the SAM2 output; if SAM2 masks are systematically biased on real shelf scenes (e.g., excluding an occluded corner or including background), a wrong pose/shape estimate can score high IoU against the biased mask, so the certificate may admit poor pseudo-labels. Please add an evaluation of SAM2 mask quality (e.g., mask IoU against the available keypoint annotations or a small set of annotated masks) on a subset of Symbotic-9k, and report the precision/recall of the 2D certificate when M is a SAM2 mask, or otherwise justify why the Fig. 4a result transfers to SAM2 masks.
  2. [Section 5.2, Table 1] The real-data pose and shape errors are computed against pseudo ground truth obtained by running the authors' estimator on ground-truth keypoints. This makes the metric measure consistency with the estimator's own bias rather than absolute accuracy, and it cannot detect systematic errors that are also present in the estimator's output on GT keypoints. Please state this limitation explicitly in the main text (it is currently only implied) and, if possible, validate on simulated data with known ground truth that the pseudo-GT procedure yields unbiased error estimates, or provide an alternative 3D ground truth for a small real subset.
  3. [Section 3, Eq. (1)] The text states that a Geman-McClure robust loss is applied to ensure robustness against outliers, but Eq. (1) shows squared L2 residuals. Please clarify the actual objective function and provide the optimization details (initialization, number of iterations, learning rate, and how the robust loss is incorporated). This is necessary for reproducibility and for interpreting the residual certificate thresholds reported in Section 4.2.
  4. [Section 4.2 and 5.3] The certificate thresholds (epsilon_2D, epsilon_res, epsilon_epi, and epsilon_conf) are tuned on the annotated validation set, yet the exact values are not reported and the tuning procedure is not described. Since these thresholds are reused for the unlabeled 50k dataset (Section 5.3), please report the exact values and include a sensitivity analysis (e.g., how keypoint RMSE varies with threshold choices) to demonstrate that the results are not overly dependent on specific threshold settings.
minor comments (6)
  1. [Figure 4b] The y-axis label 'Counts' with negative values is confusing; please clarify whether this is a signed count, a log-ratio, or something else.
  2. [Section 4.3, Eq. (4)] The caption mentions the discrepancy between rectified pseudo-label keypoints and ground truth keypoints, but Eq. (4) is defined only on predicted keypoints; please rephrase to avoid implying ground truth is used in the certificate.
  3. [Section 5.1] The phrase 'achieving accuracy more than 10 times higher' is vague; please specify the exact comparison, e.g., position error reduced from 0.584 m to 0.038 m.
  4. [Table 1] Add a footnote to Table 1 explaining that real-data pose/shape metrics use pseudo ground truth derived from the estimator on GT keypoints, as raised in the major comments.
  5. [Section 3, after Eq. (3)] Define \bar{T} immediately after Eq. (3) for clarity; currently it is introduced after the equation.
  6. [References] Reference [22] is Mask R-CNN; Keypoint-RCNN is a specialized variant, so please cite a more specific keypoint detection reference if one exists, or clarify that the keypoint head of Mask R-CNN is used.

Circularity Check

1 steps flagged · score 4.0 of 10

Real-data pose/shape metrics are scored against pseudo ground truth generated by the same estimator; keypoint-level self-supervision gains are independently benchmarked.

  1. other [Section 5.2, 'Results and Insights', Table 1]
    "Since ground truth pose and shape are unavailable for the real dataset, we generate pseudo ground truth by running our pose and shape estimator on ground truth keypoints."

    The reference pose/shape for all real-data pose/shape numbers is generated by running the paper's own PnP estimator (Eq. 1) on ground-truth keypoints, so both the evaluated model's output (from predicted keypoints) and the reference target are produced by the same geometric estimator. Any bias of the cuboid assumption, the Geman-McClure robust loss, or the PnP optimization is present on both sides, so the reported APE/ARE/ASE values measure self-consistency between predicted and annotated keypoints under the estimator rather than error against an independently measured pose. The improvement from 2.080 m to 0.134 m therefore chiefly reflects keypoint accuracy gains, which are independently supported by Fig.

full rationale

The core self-training claim has substantial independent support: the keypoint comparisons in Fig. 6b use annotated test keypoints (Symbotic-9k test split and the synthetic benchmark), and BOSS-SAM2's keypoint accuracy improvements are measured against human/GT annotations, not against the model's own outputs. The synthetic benchmark additionally uses ground-truth poses and shapes. The certificate thresholds are tuned on annotated validation curves and then applied to held-out data, which is standard practice rather than circularity. The main circular element is narrower: the real-data pose/shape metrics in Table 1 are evaluated against pseudo ground truth produced by the same estimator from GT keypoints, so those specific numbers are self-referential and should not be read as independent pose/shape validation. A separate, non-circular support gap is that the 2D certificate is validated with GT masks (Fig. 4a) but deployed with SAM2 masks; the certificate's precision on real SAM2 masks is not measured, which is an empirical robustness concern. Self-citations to the correct-and-certify paradigm [1-5] are not load-bearing because the certificates here are empirically validated and the central SSL result is benchmarked on independent keypoint annotations.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The pipeline does not introduce new physical entities or mediators. It relies on standard computer vision components, with the main free parameters being the certificate thresholds, which are hand-tuned or fitted to validation data and are load-bearing for pseudo-label quality.

free parameters (4)
  • Keypoint confidence threshold epsilon_conf = not reported
    Discards low-confidence keypoint detections; value is chosen by hand and never stated in the paper.
  • IoU certificate threshold epsilon_2D = not reported
    Threshold for the 2D certificate in Eq. 2; tuned on annotated validation data.
  • Residual certificate threshold epsilon_res = about 42 pixels (empirical)
    Hybrid predicted/reprojected keypoint selection threshold; chosen empirically from the validation curves in Fig. 4b.
  • Epipolar certificate threshold epsilon_epi = not reported
    Threshold for the epipolar certificate in Eq. 4; chosen by hand from validation trends.
assumptions (4)
  • domain assumption The object is a box parameterized by a cuboid with three side lengths and pose T.
    Introduced in Section 3; the entire PnP formulation assumes a cuboid model.
  • domain assumption SAM2 segmentation masks are accurate enough on industrial scenes to serve as pseudo-ground truth for the IoU certificate.
    Required in Section 3 and Section 4.4; validation uses GT masks rather than SAM2 masks on real warehouse images.
  • domain assumption Keypoint-RCNN trained on simulation provides sufficiently good initial keypoints for self-training to converge.
    Pre-training requirement stated in Section 3; no sensitivity study to initial model quality is provided.
  • ad hoc to paper Certificate thresholds chosen on annotated data transfer to unlabeled warehouse data.
    Section 4 derives thresholds from annotated curves and Section 5 applies the same thresholds during self-training without re-validation on the unlabeled 50k set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Box Pose and Shape Estimation and Domain Adaptation for Large-Scale Warehouse Automation." pith.science (2026). https://pith.science/paper/HHSDABRQ

@misc{pith2026250700984,
  author       = {Pith},
  title        = {Pith review of: Box Pose and Shape Estimation and Domain Adaptation for Large-Scale Warehouse Automation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HHSDABRQ}},
  note         = {Machine review of arXiv:2507.00984}
}
read the original abstract

Modern warehouse automation systems rely on fleets of intelligent robots that generate vast amounts of data -- most of which remains unannotated. This paper develops a self-supervised domain adaptation pipeline that leverages real-world, unlabeled data to improve perception models without requiring manual annotations. Our work focuses specifically on estimating the pose and shape of boxes and presents a correct-and-certify pipeline for self-supervised box pose and shape estimation. We extensively evaluate our approach across a range of simulated and real industrial settings, including adaptation to a large-scale real-world dataset of 50,000 images. The self-supervised model significantly outperforms models trained solely in simulation and shows substantial improvements over a zero-shot 3D bounding box estimation baseline.

Figures

Figures reproduced from arXiv: 2507.00984 by the authors.

Figure 1
Figure 1. (left) Robot in a Symbotic warehouse picking up a box from a shelf. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed pipeline. We take stereo images as input and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Analysis of SAM2 sampling strategies. (a) Cumulative distribution of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a)Validation of the 2D Certificate (i.e., Equation (2)). The x-axis is the IoU score in the OC2D certificate. The y-axis is the RMSE error of the pseudo￾labels averaged across one image sample. We use ground-truth segmentation for IoU calculation. Pseudo-labels with I…
Figure 5
Figure 5. Figure 5: Sample images from the simulated and real datasets used in the experi [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of keypoint detection performance in Sim2Sim (left) and [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Picasso: Holistic Scene Reconstruction with Physics-Constrained Sampling

    cs.CV 2026-02 unverdicted novelty 6.0 of 10

    Picasso produces multi-object scene reconstructions that are both geometrically accurate and physically plausible by using physics-constrained rejection sampling over an inferred contact graph, outperforming prior met...

  2. Hydra++: Real-Time Hierarchical 3D Scene Graph Construction With Object-Level Shape Estimation

    cs.CV 2026-07 conditional novelty 5.0 of 10

    Hydra++ integrates category-agnostic shape estimators (CRISP/SAM3D) plus a reprojection-mask check into real-time hierarchical scene graphs and improves object/scene reconstruction indoors and outdoors.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages · cited by 2 Pith papers

  1. [1]

    Self-supervised geometric per- ception,

    H. Yang, W. Dong, L. Carlone, and V. Koltun, “Self-supervised geometric per- ception,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14350–14361

    work page 2021

  2. [2]

    A correct-and-certify approach to self-supervise object pose estimators via ensemble self-training,

    J. Shi, R. Talak, D. Maggio, and L. Carlone, “A correct-and-certify approach to self-supervise object pose estimators via ensemble self-training,” arXiv preprint arXiv:2302.06019, 2023

    arXiv 2023

  3. [3]

    Certifiable object pose estimation: Founda- tions, learning models, and self-training,

    R. Talak, L. R. Peng, and L. Carlone, “Certifiable object pose estimation: Founda- tions, learning models, and self-training,”IEEE Transactions on Robotics, vol. 39, no. 4, pp. 2805–2824, 2023

    work page 2023

  4. [4]

    CRISP: Object Pose and Shape Estimation with Test-Time Adaptation

    J. Shi, R. Talak, H. Zhang, D. Jin, and L. Carlone, “Crisp: Object pose and shape estimation with test-time adaptation,”arXiv preprint arXiv:2412.01052 , 2024

    work page Pith review arXiv 2024

  5. [5]

    Test-time certifiable self-supervision to bridge the sim2real gap in event-based satellite pose estimation,

    M. Jawaid, R. Talak, Y. Latif, L. Carlone, and T.-J. Chin, “Test-time certifiable self-supervision to bridge the sim2real gap in event-based satellite pose estimation,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 4534–4541

    work page 2024

  6. [6]

    6-dof object pose from semantic keypoints,

    G. Pavlakos, X. Zhou, A. Chan, K. G. Derpanis, and K. Daniilidis, “6-dof object pose from semantic keypoints,” in2017 IEEE international conference on robotics and automation (ICRA) . IEEE, 2017, pp. 2011–2018

    work page 2017

  7. [7]

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

    H. Wang, S. Sridhar, J. Huang, J. Valentin, S. Song, and L. J. Guibas, “Normalized object coordinate space for category-level 6d object pose and size estimation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, 2019, pp. 2642–2651

    work page 2019

  8. [8]

    Shape prior deformation for categorical 6d object pose and size estimation,

    M. Tian, M. H. Ang Jr, and G. H. Lee, “Shape prior deformation for categorical 6d object pose and size estimation,” inEuropean Conference on Computer Vision . Springer, 2020, pp. 530–546

    work page 2020

Show all 27 references
  1. [9]

    Category-level 6d object pose estimation in the wild: A semi- supervised learning approach and a new dataset,

    Y. Fu and X. Wang, “Category-level 6d object pose estimation in the wild: A semi- supervised learning approach and a new dataset,”Advances in Neural Information Processing Systems, vol. 35, pp. 27469–27483, 2022. 12 Xihang Yu et al

    2022

  2. [10]

    Pallet pose estimation based on front face shot,

    N. Kai, H. Yoshida, and T. Shibata, “Pallet pose estimation based on front face shot,” IEEE Access, 2025

    2025

  3. [11]

    Omni3D: A large benchmark and model for 3D object detection in the wild,

    G. Brazil, A. Kumar, J. Straub, N. Ravi, J. Johnson, and G. Gkioxari, “Omni3D: A large benchmark and model for 3D object detection in the wild,” in CVPR. Vancouver, Canada: IEEE, June 2023

    2023

  4. [12]

    Unsupervised domain adaptation through self-supervision,

    Y. Sun, E. Tzeng, T. Darrell, and A. A. Efros, “Unsupervised domain adaptation through self-supervision,” arXiv preprint arXiv:1909.11825 , 2019

    1909 arXiv

  5. [13]

    Test-time training with self-supervision for generalization under distribution shifts,

    Y. Sun, X. Wang, Z. Liu, J. Miller, A. Efros, and M. Hardt, “Test-time training with self-supervision for generalization under distribution shifts,” inInternational conference on machine learning . PMLR, 2020, pp. 9229–9248

    2020

  6. [14]

    Self-supervised policy adaptation during deployment,

    N. Hansen, R. Jangir, Y. Sun, G. Alenyà, P. Abbeel, A. A. Efros, L. Pinto, and X. Wang, “Self-supervised policy adaptation during deployment,”arXiv preprint arXiv:2007.04309, 2020

    2007 arXiv

  7. [15]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,”arXiv preprint arXiv:2006.10726 , 2020

    2006 arXiv

  8. [16]

    Test time adaptation via con- jugate pseudo-labels,

    S. Goyal, M. Sun, A. Raghunathan, and J. Z. Kolter, “Test time adaptation via con- jugate pseudo-labels,” Advances in Neural Information Processing Systems, vol. 35, pp. 6204–6218, 2022

    2022

  9. [17]

    Memo: Test time robustness via adaptation and augmentation,

    M. Zhang, S. Levine, and C. Finn, “Memo: Test time robustness via adaptation and augmentation,” Advances in neural information processing systems , vol. 35, pp. 38629–38642, 2022

    2022

  10. [18]

    Embodied ac- tive domain adaptation for semantic segmentation via informative path planning,

    R. Zurbrügg, H. Blum, C. Cadena, R. Siegwart, and L. Schmid, “Embodied ac- tive domain adaptation for semantic segmentation via informative path planning,” IEEE Robotics and Automation Letters , vol. 7, no. 4, pp. 8691–8698, 2022

    2022

  11. [19]

    Symmetry and uncertainty-aware object slam for 6dof object pose estimation,

    N. Merrill, Y. Guo, X. Zuo, X. Huang, S. Leutenegger, X. Peng, L. Ren, and G. Huang, “Symmetry and uncertainty-aware object slam for 6dof object pose estimation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14901–14910

    2022

  12. [20]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollár, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,” arXiv preprint arXiv:2408.007...

    2024 arXiv

  13. [21]

    A general and adaptive robust loss function,

    J. T. Barron, “A general and adaptive robust loss function,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 4331–4339

    2019

  14. [22]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask r-cnn,” inProceedings of the IEEE international conference on computer vision , 2017, pp. 2961–2969

    2017

  15. [23]

    Pytorch: An imperative style, high-performance deep learning library,

    A.Paszke,S.Gross,F.Massa,A.Lerer,J.Bradbury,G.Chanan,T.Killeen,Z.Lin, N. Gimelshein, L. Antigaet al., “Pytorch: An imperative style, high-performance deep learning library,”Advances in neural information processing systems , vol. 32, 2019

    2019

  16. [24]

    Pharr, W

    M. Pharr, W. Jakob, and G. Humphreys,Physically based rendering: From theory to implementation. MIT Press, 2023

    2023

  17. [25]

    Ellipsdf: Joint object pose and shape optimization with a bi-level ellipsoid and signed distance function descrip- tion,

    M. Shan, Q. Feng, Y.-Y. Jau, and N. Atanasov, “Ellipsdf: Joint object pose and shape optimization with a bi-level ellipsoid and signed distance function descrip- tion,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5946–5955

    2021

  18. [26]

    Autosdf: Shape priors for 3d completion, reconstruction and generation,

    P. Mittal, Y.-C. Cheng, M. Singh, and S. Tulsiani, “Autosdf: Shape priors for 3d completion, reconstruction and generation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 306–315

    2022

  19. [27]

    Openshape: Scaling up 3d shape representation towards open-world understand- ing,

    M. Liu, R. Shi, K. Kuang, Y. Zhu, X. Li, S. Han, H. Cai, F. Porikli, and H. Su, “Openshape: Scaling up 3d shape representation towards open-world understand- ing,” Advances in neural information processing systems, vol. 36, pp. 44860–44879, 2023

    2023

Pith tools

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