Pith. sign in

REVIEW 3 major objections 6 minor 47 references

HOGSA: Bimanual Hand-Object Interaction Understanding with 3D Gaussian Splatting Based Data Augmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that augmenting bimanual hand-object interaction datasets with mesh-based 3D Gaussian Splatting renders—sharpened by super-resolution and diversified by a contact-consistency pose optimizer—improves downstream pose and…

desk verdict HOGSA is a useful engineering contribution with a real speed advantage, but the paper overclaims the mechanism: the gains could partly come from simply having more data, and there are no error bars or matched-compute controls. read the letter →

arxiv 2501.02845 v1 pith:ZMAKKTZZ submitted 2025-01-06 cs.CV

Wentian Qu , Jiahe Li , Jian Cheng , Jian Shi , Chenyu Meng , Cuixia Ma , Hongan Wang , Xiaoming Deng
show 1 more author
Yinda Zhang
This is my paper · ORCID
classification cs.CV
keywords bimanualhand-objectinteractiondataaugmentation3DGaussianSplattingposeoptimizationsuper-resolutioncontactestimationmotionreconstructionsynthetictraining
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

Understanding what two hands are doing with an object—estimating hand poses, contact regions, and relative motion—needs large annotated video datasets, but bimanual interactions are severely occluded and hard to annotate. This paper tries to remove that bottleneck by automatically generating photorealistic training images from the datasets themselves. It builds a mesh-based 3D Gaussian Splatting model of each hand and the object per sequence, optimises new bimanual poses that satisfy contact consistency, and runs renders through a super-resolution network. Combining the generated images with the original data improves standard baselines on the Arctic and H2O benchmarks, reducing contact deviation, motion deviation, and hand pose error on held-out subjects. The paper's claim is that high-quality synthetic two-hand interaction data can be produced without manual 3D scanning, and that this data transfers to real-world interaction understanding.

What carries the argument

The load-bearing object is the mesh-based Hand-Object Gaussian Splatting (HOGS) model: each Gaussian kernel is anchored to a mesh face through barycentric coordinates $\beta$, so the kernel position stays tied to the MANO-HD hand mesh or the object mesh, and hand motion is applied via skinning weights. Around this core sit two modules: the Pose Optimization Module (POM), which extends GraspTTA's single-hand contact-consistency optimisation to two hands by perturbing initial poses and minimising a contact-consistency loss $\mathcal{L}_C$ plus hand-centric and penetration losses; and the Super-Resolution Module (SRM), a StyleUNet GAN that refines coarse 3DGS renders. The mesh anchoring prevents the Gaussian points from straying off the surface, which the authors identify as the main source of artifacts in vanilla 3DGS augmentation; POM provides pose diversity; SRM provides texture realism.

What would settle it

Render POM-generated poses with hand meshes and measure the maximum interpenetration depth between each hand mesh and the object mesh, comparing POM outputs against the original dataset poses and against random perturbations; if POM poses penetrate as deeply as random ones, the contact-consistency loss is not enforcing physical plausibility and the reported baseline gains must come from viewpoint and pose diversity alone. Alternatively, replace the ContactNet contact maps in the loss with random maps of the same statistics and check whether the downstream improvements vanish.

Watch

Extended reading notes

Core claim

The central discovery, stated in the paper's terms, is that a data augmentation framework built on mesh-constrained 3D Gaussian Splatting can automatically expand a bimanual hand-object interaction dataset to roughly double its size with photorealistic images of novel poses and viewpoints, and that fine-tuning interaction-understanding baselines on the combined data improves their accuracy. On Arctic, fine-tuning ArcticNet-SF with HOGSA reduces contact deviation from 41.35 to 35.23 mm, reduces hand mean per-joint position error from 23.01 to 20.96 mm, and raises grasp success rate from 71.77 to 77.85 percent; on H2O it raises success rate from 39.80 to 45.27 percent. The interaction-field baseline InterField-SF also improves on both benchmarks. The authors attribute the gains to three components: mesh-anchored Gaussians that keep geometry sharp, a pose optimizer that expands pose distribution while penalising penetration and enforcing contact consistency, and a super-resolution module that closes the gap between rendered and real images.

Load-bearing premise

The pipeline assumes that ContactNet, a contact-prior network trained on single-hand grasps, judges bimanual contact quality accurately enough to generate physically plausible two-hand poses, and the paper does not directly measure grasp quality, only downstream metrics.

Editorial extensions

If this is right

  • Fine-tuning with HOGSA-augmented data lowers contact deviation, motion deviation, and hand pose error for the motion-reconstruction baseline on both Arctic and H2O.
  • The augmentation nearly doubles the available training data (113% on Arctic, 175% on H2O) and renders each new image in about 0.06 seconds, making large-scale synthetic fine-tuning practical.
  • The interaction-field baseline also improves on both benchmarks, showing the method helps tasks beyond pose reconstruction.
  • The same HOGS models support novel-view synthesis and pose transfer to other subjects, so the framework is reusable beyond augmentation.

Reading between the lines

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

  • Inference: Because HOGS is built per sequence from plain video frames, the pipeline should transfer to other bimanual video datasets (such as egocentric interaction footage) without manual 3D scanning; a direct check would be applying HOGSA to an egocentric two-hand dataset and measuring baseline gains.
  • Inference: The method would likely benefit from a bimanual-specific contact prior trained on Arctic/H2O contact annotations; the paper reuses a single-hand prior, so a two-hand prior could generate a larger space of plausible poses and push gains further.
  • Inference: The ablation suggests rendering quality and pose diversity contribute independently; a sharper test would hold pose distribution fixed and vary only the SRM, isolating how much of the gain is realism versus pose coverage.
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

3 major / 6 minor

Summary. The paper proposes HOGSA, a 3D Gaussian Splatting-based data augmentation framework for bimanual hand-object interaction. The method builds mesh-based 3DGS models of hands and objects from existing datasets, uses a Pose Optimization Module (POM) extended from GraspTTA to synthesize new bimanual hand-object poses, and a Super-Resolution Module (SRM) to improve rendering quality. The augmented data are combined with original training data to fine-tune two interaction-understanding baselines (ArcticNet-SF and InterField-SF) on Arctic and H2O. The paper reports consistent improvements across contact, motion, pose, and interaction-field metrics, and provides ablations showing that removing SRM or POM degrades performance.

Significance. If the reported gains are robust, the paper offers a practical and efficient alternative to hand-crafted synthetic datasets for bimanual hand-object interaction, with a generation time of 0.06s per image, a public project page, and a systematic analysis of rendering quality and pose diversity. The evaluation uses held-out test subjects and external benchmarks, which avoids circular fitting, and the paper is transparent about the data splits. The main weaknesses are that the empirical improvements are confounded with the increased training set size and optimizer steps, and that the physical plausibility of POM-generated poses is never directly validated; these issues currently limit the strength of the central claim that the proposed modules specifically drive the improvement.

major comments (3)
  1. [§4.2, Table 1, Appendix §2.1] The core comparison between the baseline and 'Baseline+HOGSA' holds the fine-tuning epoch count fixed (15 epochs, Appendix §2.1) while the augmented datasets add 113% (Arctic) and 175% (H2O) more images. The augmented runs therefore process approximately 2.1 and 2.75 times as many optimizer steps per epoch as the baseline runs. Since no control is reported that matches compute or data volume (e.g., training the baseline on duplicated original data, on the same total number of iterations, or on an equally enlarged randomly-jittered dataset), Table 1's improvements cannot be uniquely attributed to HOGSA's specific modules. The same confound applies to Table 4: it is not stated whether the 'w/o SRM', 'w/o POM', and 'Full' rows use identical augmented-image counts and identical total iterations, so the differences (e.g., CDev 35.23 vs 36.85/37.67) may reflect data scale rather than the SRM or POM modules. Please add matched-compute or matched-data-volume controls and report the per-condition data sizes and iteration counts.
  2. [§3.3, Table 4] The paper repeatedly describes POM-generated poses as 'feasible' and 'physically plausible', but this is never directly evaluated. No metric on the generated poses is reported (e.g., penetration depth, contact accuracy, or success rate of the optimization), and the only evidence is downstream baseline improvement, which is confounded with data scale as discussed above. In addition, ContactNet is borrowed from single-hand GraspTTA and applied to bimanual interactions without validation that its contact prior transfers. Please provide a direct assessment of generated grasp quality, such as penetration volume, hand-object contact deviation, or a perceptual study, and report it alongside the ablation.
  3. [§4.2, Tables 1–4] All quantitative results are reported as single runs without error bars or significance tests. Given that several improvements are small (e.g., ACC_o drops from 6.05 to 6.02 on H2O in Table 1, or ACC from 1.84 to 1.82 in Table 2), it is not clear that these differences are statistically meaningful. Please report the mean and standard deviation over at least three random seeds, and where appropriate a paired test, to support the claim that HOGSA 'improves the performance' on all metrics.
minor comments (6)
  1. [§2] There is a typo in the Related Work section: 'augumented' should be 'augmented'.
  2. [§4.1] The metric name 'LIPIPS' in Section 4.1 and Table 3 appears to be a typo for 'LPIPS'.
  3. [§4.2] The timing comparison of 0.06s per image versus HOIDiffusion's 4.5s omits the up-front cost of per-sequence HOGS and SRM training; please state the total end-to-end cost of the augmentation pipeline.
  4. [Table 2] The notation 'w' and 'w/o' in Table 2 is ambiguous; consider using 'with' and 'without' in the caption or a clear legend.
  5. [Figure 7] The T-SNE visualization is qualitative; a quantitative measure of pose-distribution coverage would make the diversity claim more precise.
  6. [§1 (Contributions)] The contribution list claims 'state-of-the-art results on H2O and Arctic benchmarks', but the experiments only compare with HOIDiffusion and a 3DGS variant; no comparison with other published methods on these benchmarks is provided. Please either add comparisons or revise the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HOGSA is validated on held-out subjects and external baselines, with no prediction fitted to the test data.

full rationale

The paper's central claim is an empirical data-augmentation result: HOGSA-generated images are produced from training-subject sequences only, combined with the original training set, and used to fine-tune standard baselines whose metrics are reported on held-out subjects (Arctic s03/s05 and H2O subject4). The HOGS, POM, and SRM modules are trained on the training split and rely on external components (MANO-HD, GaMeS-style mesh Gaussian splatting, GraspTTA's ContactNet, and a StyleAvatar-inspired super-resolution network), none of which encode the target test outputs. The self-citations, including HO-NeRF (Qu et al. 2023) and the rendering-dataset paper (Deng et al. 2021), appear as related-work motivation and are not used to justify the claimed improvement, so they are not load-bearing. No equation or fitted parameter makes the measured improvement equal to an input by construction. The main weakness is that augmented runs see substantially more images per epoch than the unaugmented controls, which is a compute-matching confound rather than a circularity; the evaluation itself remains a held-out, externally benchmarked comparison.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method is an assembly of existing components (mesh-based 3DGS, GraspTTA's ContactNet, StyleAvatar-style GAN) with a new bimanual pose optimization. It does not introduce new physical entities, but it relies on several hand-tuned parameters and external priors; the most fragile is the transferability of the single-hand contact prior to bimanual scenarios.

free parameters (3)
  • Loss hyperparameters = λSSIM=0.2, λR=0.5, λC=1, λH=1, λP=17, λ1=5, λVGG=0.03
    Hand-tuned weights for HOGS, POM, and SRM losses; they control the balance between photo-realism, geometric plausibility, and pose diversity in the generated data.
  • POM perturbation ranges = Random rotation [0,20 degrees], 5 percent distance translation, [0,6 cm] position offset
    Chosen heuristically to expand pose distribution; no principled selection or sensitivity analysis is provided.
  • Gaussian kernels per mesh face k = 2
    Design choice in mesh-based 3DGS defining model capacity; may affect rendering fidelity and training cost.
assumptions (4)
  • domain assumption MANO-HD hand model and its linear blend skinning accurately represent hand shape and pose for both hands across subjects.
    Used in Eq. 3 to deform hand Gaussian kernels into target poses; if inaccurate, synthetic hand images will be unrealistic.
  • domain assumption The datasets provide reliable object meshes and articulation parameters (including Arctic articulated objects) used to build object Gaussian models.
    Object model initialization and rendering assume mesh availability; errors in object shape or articulation would propagate into augmented images.
  • domain assumption ContactNet from GraspTTA can be frozen and applied to bimanual poses, providing a valid contact prior.
    POM calculates consistency loss LC between computed and predicted contact maps; the prior was trained on single-hand data and its transfer to two hands is assumed.
  • domain assumption The GAN-based super-resolution module improves texture realism without distorting geometry or pose.
    SRM is trained on synthetic-render to real-image pairs; it assumes the refinement is appearance-only and does not introduce pose bias into training data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HOGSA: Bimanual Hand-Object Interaction Understanding with 3D Gaussian Splatting Based Data Augmentation." pith.science (2026). https://pith.science/paper/ZMAKKTZZ

@misc{pith2026250102845,
  author       = {Pith},
  title        = {Pith review of: HOGSA: Bimanual Hand-Object Interaction Understanding with 3D Gaussian Splatting Based Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMAKKTZZ}},
  note         = {Machine review of arXiv:2501.02845}
}
read the original abstract

Understanding of bimanual hand-object interaction plays an important role in robotics and virtual reality. However, due to significant occlusions between hands and object as well as the high degree-of-freedom motions, it is challenging to collect and annotate a high-quality, large-scale dataset, which prevents further improvement of bimanual hand-object interaction-related baselines. In this work, we propose a new 3D Gaussian Splatting based data augmentation framework for bimanual hand-object interaction, which is capable of augmenting existing dataset to large-scale photorealistic data with various hand-object pose and viewpoints. First, we use mesh-based 3DGS to model objects and hands, and to deal with the rendering blur problem due to multi-resolution input images used, we design a super-resolution module. Second, we extend the single hand grasping pose optimization module for the bimanual hand object to generate various poses of bimanual hand-object interaction, which can significantly expand the pose distribution of the dataset. Third, we conduct an analysis for the impact of different aspects of the proposed data augmentation on the understanding of the bimanual hand-object interaction. We perform our data augmentation on two benchmarks, H2O and Arctic, and verify that our method can improve the performance of the baselines.

Figures

Figures reproduced from arXiv: 2501.02845 by the authors.

Figure 1
Figure 1. We propose a new 3DGS-based data augmentation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our data augmentation framework for bimanual hand-object interaction. Based on the original dataset, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of our HOGSA, which contains diverse [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The augmented data we used to train the baseline. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of our data augmentation method HOGSA on the baseline. After optimization, the model can cover [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on SRM. After using SRM, the realism of the image rendered by HOGS is significantly improved, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: We used the joints of both hands to perform T [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Gaussian kernel defined on mesh-based 3DGS. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Comparison of interaction field estimation. We find that after we optimize the baseline with our data augmentation [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 12
Figure 12. Figure 12: We find that mesh-based methods can preserve [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 10
Figure 10. Figure 10: The diversity of interactive poses after our POM [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Comparison of consistent motion reconstruction. After optimizing the baseline using our data augmentation, the [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Comparison of rendering quality. Our mesh-based method captures more realistic textures, and our SRM can be [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Our method can support applications such as novel view synthesis and pose transfer. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 33 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P

    Barron, J. T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P. P. 2021. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 5855--5864

    2021

  4. [4]

    D.; Kemp, C

    Brahmbhatt, S.; Tang, C.; Twigg, C. D.; Kemp, C. C.; and Hays, J. 2020. ContactPose: A dataset of grasps with object contact and hand pose. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XIII 16, 361--378. Springer

    work page 2020

  5. [5]

    Cao, Z.; Radosavovic, I.; Kanazawa, A.; and Malik, J. 2021. Reconstructing hand-object interactions in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 12417--12426

    work page 2021

  6. [6]

    Chen, X.; Wang, B.; and Shum, H.-Y. 2023. Hand avatar: Free-pose hand animation and rendering from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8683--8693

    work page 2023

  7. [7]

    Chen, Z.; Chen, S.; Schmid, C.; and Laptev, I. 2023. gsdf: Geometry-driven signed distance functions for 3d hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12890--12900

    work page 2023

  8. [8]

    Corona, E.; Pumarola, A.; Alenya, G.; Moreno-Noguer, F.; and Rogez, G. 2020. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5031--5041

    work page 2020

Show all 47 references
  1. [9]

    Deng, X.; Zhang, Y.; Shi, J.; Zhu, Y.; Cheng, D.; Zuo, D.; Cui, Z.; Tan, P.; Chang, L.; and Wang, H. 2021. Hand pose understanding with large-scale photo-realistic rendering dataset. IEEE Transactions on Image Processing, 30: 4275--4290

    2021

  2. [10]

    E.; Chen, X.; Kocabas, M.; Black, M

    Fan, Z.; Parelli, M.; Kadoglou, M. E.; Chen, X.; Kocabas, M.; Black, M. J.; and Hilliges, O. 2024. HOLD: Category-agnostic 3d reconstruction of interacting hands and objects from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 494--504

    2024

  3. [11]

    J.; and Hilliges, O

    Fan, Z.; Taheri, O.; Tzionas, D.; Kocabas, M.; Kaufmann, M.; Black, M. J.; and Hilliges, O. 2023. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12943--12954

    2023

  4. [12]

    Feldmann, C.; Siegenheim, N.; Hars, N.; Rabuzin, L.; Ertugrul, M.; Wolfart, L.; Pollefeys, M.; Bauer, Z.; and Oswald, M. R. 2024. NeRFmentation: NeRF-based Augmentation for Monocular Depth Estimation. arXiv preprint arXiv:2401.03771

    2024 arXiv

  5. [13]

    Gao, D.; Xiu, Y.; Li, K.; Yang, L.; Wang, F.; Zhang, P.; Zhang, B.; Lu, C.; and Tan, P. 2022. DART: Articulated hand model with diverse accessories and rich textures. Advances in Neural Information Processing Systems, 35: 37055--37067

    2022

  6. [14]

    D.; Vo, M.; Brahmbhatt, S.; and Kemp, C

    Grady, P.; Tang, C.; Twigg, C. D.; Vo, M.; Brahmbhatt, S.; and Kemp, C. C. 2021. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1471--1481

    2021

  7. [15]

    Hu, H.; Wang, W.; Zhou, W.; and Li, H. 2022. Hand-object interaction image generation. Advances in Neural Information Processing Systems, 35: 23805--23817

    2022

  8. [16]

    Jian, J.; Liu, X.; Li, M.; Hu, R.; and Liu, J. 2023. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 14713--14724

    2023

  9. [17]

    Jiang, H.; Liu, S.; Wang, J.; and Wang, X. 2021. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, 11107--11116

    2021

  10. [18]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  11. [19]

    Kwon, T.; Tekin, B.; St \"u hmer, J.; Bogo, F.; and Pollefeys, M. 2021. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 10138--10148

    2021

  12. [20]

    Li, K.; Yang, L.; Zhen, H.; Lin, Z.; Zhan, X.; Zhong, L.; Xu, J.; Wu, K.; and Lu, C. 2023. Chord: Category-level hand-held object reconstruction via shape deformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9444--9454

    2023

  13. [21]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 1...

    2014

  14. [22]

    Liu, S.; Jiang, H.; Xu, J.; Liu, S.; and Wang, X. 2021. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14687--14697

    2021

  15. [23]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  16. [24]

    Mueller, F.; Bernard, F.; Sotnychenko, O.; Mehta, D.; Sridhar, S.; Casas, D.; and Theobalt, C. 2018. Ganerated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of the IEEE conference on computer vision and pattern recognition, 49--59

    2018

  17. [25]

    Narasimhaswamy, S.; Nguyen, T.; and Nguyen, M. H. 2020. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing systems, 33: 7841--7851

    2020

  18. [26]

    N.; Xing, A.; Li, Z.; Chen, K.; Sharma, A.; and Sridhar, S

    Pokhariya, C.; Shah, I. N.; Xing, A.; Li, Z.; Chen, K.; Sharma, A.; and Sridhar, S. 2023. MANUS: Markerless Grasp Capture using Articulated 3D Gaussians. arXiv preprint arXiv:2312.02137

    2023 arXiv

  19. [27]

    Qi, H.; Zhao, C.; Salzmann, M.; and Mathis, A. 2024. HOISDF: Constraining 3D Hand-Object Pose Estimation with Global Signed Distance Fields: Processed data and trained models. In The 2024 IEEE/CVF Computer Vision and Pattern Recognition Conference. Zenodo

    2024

  20. [28]

    Qian, N.; Wang, J.; Mueller, F.; Bernard, F.; Golyanik, V.; and Theobalt, C. 2020. Html: A parametric hand texture model for 3d hand reconstruction and personalization. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part...

    2020

  21. [29]

    Qu, W.; Cui, Z.; Zhang, Y.; Meng, C.; Ma, C.; Deng, X.; and Wang, H. 2023. Novel-view synthesis and pose estimation for hand-object interaction from sparse views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 15100--15111

    2023

  22. [30]

    S.; and Ramanan, D

    Rogez, G.; Supancic, J. S.; and Ramanan, D. 2015. Understanding everyday hands in action from rgb-d images. In Proceedings of the IEEE international conference on computer vision, 3889--3897

    2015

  23. [31]

    Romero, J.; Tzionas, D.; and Black, M. J. 2022. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610

    2022 arXiv

  24. [32]

    Shan, D.; Geng, J.; Shu, M.; and Fouhey, D. F. 2020. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9869--9878

    2020

  25. [33]

    J.; and Tzionas, D

    Taheri, O.; Ghorbani, N.; Black, M. J.; and Tzionas, D. 2020. GRAB: A dataset of whole-body human grasping of objects. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part IV 16, 581--600. Springer

    2020

  26. [34]

    Tekin, B.; Bogo, F.; and Pollefeys, M. 2019. H+ o: Unified egocentric recognition of 3d hand-object poses and interactions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4511--4520

    2019

  27. [35]

    Van der Maaten, L.; and Hinton, G. 2008. Visualizing data using t-SNE. Journal of machine learning research, 9(11)

    2008

  28. [36]

    Waczy \'n ska, J.; Borycki, P.; Tadeja, S.; Tabor, J.; and Spurek, P. 2024. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459

    2024 arXiv

  29. [37]

    Wang, L.; Zhao, X.; Sun, J.; Zhang, Y.; Zhang, H.; Yu, T.; and Liu, Y. 2023. Styleavatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, 1--10

    2023

  30. [38]

    Wang, P.; Liu, L.; Liu, Y.; Theobalt, C.; Komura, T.; and Wang, W. 2021. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689

    2021 arXiv

  31. [39]

    Xu, H.; Li, H.; Wang, Y.; Liu, S.; and Fu, C.-W. 2024. HandBooster: Boosting 3D Hand-Mesh Reconstruction by Conditional Synthesis and Sampling of Hand-Object Interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10159--10169

    2024

  32. [40]

    Xu, Y.; Wan, W.; Zhang, J.; Liu, H.; Shan, Z.; Shen, H.; Wang, R.; Geng, H.; Weng, Y.; Chen, J.; et al. 2023. UniDexGrasp: Universal Robotic Dexterous Grasping via Learning Diverse Proposal Generation and Goal-Conditioned Policy. arXiv preprint arXiv:2303.00938

    2023 arXiv

  33. [41]

    Yang, L.; Li, K.; Zhan, X.; Lv, J.; Xu, W.; Li, J.; and Lu, C. 2022. Artiboost: Boosting articulated 3d hand-object pose estimation via online exploration and synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2750--2760

    2022

  34. [42]

    Yang, L.; Zhan, X.; Li, K.; Xu, W.; Zhang, J.; Li, J.; and Lu, C. 2024. Learning a contact potential field for modeling the hand-object interaction. IEEE transactions on pattern analysis and machine intelligence

    2024

  35. [43]

    Ye, Y.; Gupta, A.; and Tulsiani, S. 2022. What's in your hands? 3d reconstruction of generic objects in hands. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3895--3905

    2022

  36. [44]

    Ye, Y.; Li, X.; Gupta, A.; De Mello, S.; Birchfield, S.; Song, J.; Tulsiani, S.; and Liu, S. 2023. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22479--22489

    2023

  37. [45]

    Yu, Z.; Chen, A.; Huang, B.; Sattler, T.; and Geiger, A. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19447--19456

    2024

  38. [46]

    Zhang, C.; Jiao, G.; Di, Y.; Wang, G.; Huang, Z.; Zhang, R.; Manhardt, F.; Fu, B.; Tombari, F.; and Ji, X. 2024 a . Moho: Learning single-view hand-held object reconstruction with multi-view occlusion-aware supervision. In Proceedings of the IEEE/CVF Conference on Computer Vis...

    2024

  39. [47]

    Zhang, M.; Fu, Y.; Ding, Z.; Liu, S.; Tu, Z.; and Wang, X. 2024 b . Hoidiffusion: Generating realistic 3d hand-object interaction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8521--8531

    2024

Pith tools

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