Pith. sign in

REVIEW 3 major objections 3 minor 60 references

Coarse hand-object reconstructions can be corrected by a gravity-field diffusion bridge that pulls the hand surface onto the object while avoiding intersections and gaps.

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 · deepseek-v4-flash

2026-08-05 11:06 UTC pith:SQCX4E3O

load-bearing objection Plausible new optimization combo, but SHOWMe SOTA is undercut by the authors' own ground-truth refinement and metrics that the method directly minimizes. the 3 major comments →

arxiv 2509.03114 v1 pith:SQCX4E3O submitted 2025-09-03 cs.CV

Towards Realistic Hand-Object Interaction with Gravity-Field Based Diffusion Bridge

Miao Xu , Xiangyu Zhu , Xusheng Liang , Zidu Wang , Jinlin Wu , Zhen Lei This is my paper
classification cs.CV
keywords hand-object interactionhand deformationdiffusion bridgegravity potential fieldcontact region inferencelarge language modelgrasp refinementmesh interpenetration
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.

The paper tries to establish that a coarse, physically implausible hand-object reconstruction can be turned into a realistic grasp by treating contact as gravitational attraction: the object's surface is a zero-potential manifold, and a diffusion bridge attracts hand points toward it while a MANO shape prior and a normal-alignment loss keep the hand anatomically plausible. If true, this gives a post-processing module that removes mesh interpenetration, closes grasping gaps, and produces hand deformation at contact regions, without retraining upstream pose or reconstruction models. The authors also claim that language-guided contact masks, inferred by an LLM from a text instruction, improve which regions of hand and object come into contact, going beyond purely geometric nearest-point selection. They support the claim with reported best numbers on SHOWMe, GRAB, and HO3D-V3, especially for penetration volume/depth and proximity error, and with ablations showing each component contributes.

Core claim

The central claim is that hand-object interaction can be modeled as an attraction-driven stochastic process rather than learned as direct regression. GravityDB represents the object surface as a zero-potential manifold and builds a multi-scale Gaussian potential field around predicted contact regions; the field's gradient acts like gravity, pulling each hand point toward the surface from either side, which simultaneously corrects interpenetration and closes gaps. The evolution of hand points is governed by a stochastic differential equation whose drift combines this gravitational force with two regularizers, a MANO-based anatomical prior and a contact-normal alignment term, and whose diffusi

What carries the argument

The machinery is a multi-scale Gaussian potential field φ(X) whose zero-potential manifold is the object surface, with weights w_p^l set by semantic contact masks. The negative gradient −∇φ supplies a gravitational drift in a stochastic differential equation, dXt = −∇φ(Xt)dt − λ1∇L_MANOdt − λ2∇L_normaldt + g(Xt)dWt, solved by Euler–Maruyama. The gravitational term attracts hand points to predicted contact regions and repels them from inside the object; L_MANO keeps the deformation close to the MANO parametric hand model; L_normal forces contacting hand points to approach along the object surface normal; the noise term g dWt is annealed for smooth convergence. A large-language-model contact i

Load-bearing premise

The load-bearing premise is that the hand mesh ground truth used to judge deformation quality, namely the SHOWMe meshes refined by the authors with details deferred to an appendix not present in the submitted text, is an unbiased target rather than a target that already favors GravityDB's deformation model.

What would settle it

A decisive check: recompute the SHOWMe comparisons using the original un-refined hand meshes as ground truth, and also compare GravityDB outputs against an independent high-resolution 3D capture of real hands performing the same grasps. If the reported margins over GeneOH and TOCH shrink substantially under either target, the realism gains are partly a product of the evaluation target rather than the deformation itself.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • GravityDB can be appended to any coarse hand-object estimate, whether from pose estimation, reconstruction, or tracking, without retraining the upstream model, since its inputs are only a hand point cloud, an object point cloud, and an optional text instruction.
  • Penetration and gap artifacts should largely disappear even when the initial pose is imperfect, because the zero-potential manifold pulls hand points to the surface from either side.
  • Hand outputs remain near the MANO manifold, so downstream applications that assume parametric hand models can consume the refined meshes without anatomical correction.
  • Contact regions become controllable by semantics: the same object can be grasped differently depending on the instruction, with contact masks guiding where the attraction field is strong.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the same attraction-based SDE is not hand-specific; with a different anatomical prior or none, it could refine body-object or foot-ground contacts, though the paper only demonstrates hands.
  • Because the object is treated as a rigid zero-potential surface, the formulation would need a reworked moving potential field to handle deformable or articulated objects; the paper does not address that regime.
  • The reported deformation-quality comparisons rest on SHOWMe ground-truth meshes that the authors refined themselves, with the procedure deferred to an appendix not included in the submitted text; an external validation of that refinement would determine whether the realism gains are measuring genuine deformation or largely reflecting the refinement convention.
  • The evaluation selects frames where the hand is already grasping and uses ground-truth object meshes, so it tests refinement of existing contact rather than full grasp synthesis from a non-contact start.

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 / 3 minor

Summary. The paper proposes GravityDB, a post-processing module that deforms a MANO-based hand point cloud toward an object surface using a multi-scale gravity field embedded in an SDE, with optional LLM-inferred contact regions. The method is evaluated as a HOI denoiser and reconstruction post-processor on SHOWMe, GRAB, and HO3D, reporting large gains on penetration (IV, PD) and proximity (PE) metrics, as well as hand-shape errors (MPJPE, MPVPE). The core claim is that GravityDB generates physically plausible, deformation-aware hand-object interactions without interpenetration or grasping gaps.

Significance. If validated, GravityDB would be a simple and transparent post-processing tool for converting coarse hand-object reconstructions into contact-stable, deformed hand meshes. The use of LLM-predicted contact regions is an interesting semantic addition. The paper provides an explicit SDE formulation, ablations for each loss term, and a comparison with several baselines. However, the current evaluation has a load-bearing flaw: the SHOWMe ground-truth hand meshes were refined by the authors, and the refinement procedure is deferred to a missing Appendix D. Because the method itself optimizes toward the object surface with MANO and normal constraints, the refined ground truth may encode the same deformation manifold GravityDB is designed to reach. The external GRAB/HO3D results provide partial support, but the paper's claim of consistent state-of-the-art performance is not supported by the reported numbers.

major comments (3)
  1. [Section 4.4 / Appendix D] The SHOWMe evaluation is not controlled. The paper states that "the hand mesh ground truth in the SHOWMe dataset has been refined by us to better reflect realistic deformation," but the promised Appendix D describing this refinement is not present in the submitted text. This is load-bearing because Tables 1 and 2 use SHOWMe as the primary dataset for the headline SOTA claims on MPJPE, MPVPE, IV, PD, and PE. Since GravityDB is a post-processing optimizer whose drift terms attract hand points toward the object surface, enforce MANO plausibility, and align contact normals, the author-refined ground truth may lie precisely on the manifold that GravityDB is designed to produce. Baselines such as TOCH, GeneOH, and HoMan are not retrained or re-tuned on this modified ground truth, so the comparison is not controlled. The authors must either include the full refinement procedure and demonstrate
  2. [Section 3.3, Eq. (7)-(8), Tables 1-2] The large reductions in IV, PD, and PE are partly by construction. The SDE update in Eq. (8) is a gradient-descent-like process on a potential that attracts hand points to the object surface, plus an explicit normal-alignment loss Lnormal, plus a MANO prior. These are the same geometric quantities that IV, PD, and PE measure: distance to the object surface and interpenetration volume/depth. Any method that simply pulls hand points toward the object surface would be expected to improve these metrics. This does not invalidate the method, but it means the metric improvements are not evidence of physically realistic grasping. To support the central claim, the authors should add a straightforward deterministic baseline (e.g., surface projection or ICP-like attraction without diffusion, and without LLM guidance) and show that GravityDB outperforms it. They should also report contact metrics on
  3. [Section 4.3, Table 1] The text claim that the method "consistently outperforms state-of-the-art baselines across key metrics" is internally overstated. In Table 1, on GRAB, Ours has MPVPE 10.05 mm versus GeneOH's 9.23 mm; on HO3D, Ours has MPVPE 12.07 mm versus GeneOH's 11.31 mm. Thus GravityDB does not consistently improve hand-shape error on these external datasets. The paper should qualify the conclusion to say that GravityDB improves penetration and proximity metrics while hand-shape accuracy is dataset-dependent. Additionally, Section 4.3 refers to HOMC (5.29 mm²) as a key metric, but HOMC does not appear in Table 1 or elsewhere in the experiments; please add the HOMC column/table or remove the claim.
minor comments (3)
  1. [Throughout] There are several typographical issues: "V olume" in Table 1, "addtitional" in Section 3.4, "instrument" in Section 4.1 (should be "object"), and "ShOWMe" capitalization inconsistency. Please proofread the manuscript.
  2. [Section 4.5] The ablation study is helpful, but the text in the paragraph after Table 3 appears to have a typo: "from 17.43 mm to 8.23 mm" should presumably read "from 8.23 mm to 17.43 mm" when describing the effect of removing LMANO. Please correct.
  3. [References] Reference [31] is a PhD thesis; please verify the citation format and page numbers. Also, several entries have inconsistent capitalization of titles.

Circularity Check

2 steps flagged

SHOWMe SOTA rests on author-refined ground truth and objective metrics; GRAB/HO3D provide only partial external grounding.

specific steps
  1. fitted input called prediction [Section 4.4 and Introduction (dataset construction)]
    "It is worth noting that the hand mesh ground truth in the SHOWMe dataset has been refined by us to better reflect realistic deformation. This ensures a more accurate evaluation of deformation quality and interaction metrics."

    Tables 1 and 2 use SHOWMe to claim state-of-the-art MPJPE, MPVPE, IV, PD, and PE. But the hand-mesh ground truth is not the original external dataset: it is a version 'refined by us,' and the refinement procedure is deferred to an Appendix D that is absent from the submitted text. GravityDB is itself a deformation procedure that attracts hand vertices to the object surface and regularizes with MANO and normal-alignment terms. If the withheld refinement used comparable contact/deformation operations, then the evaluation target lies on the same manifold that GravityDB is designed to reach, making the SHOWMe SOTA gap partly self-fulfilling. Because the refinement is unverifiable, the central SHOWMe benchmark cannot be treated as independent evidence.

  2. self definitional [Section 3.3 (termination condition) and Table 1]
    "The diffusion is terminated when one of the following conditions is satisfied: ... (iii) all contact validity criteria are met (no penetrations, all assigned contact regions are occupied). ... As shown in Table 1, our method consistently outperforms state-of-the-art baselines across key metrics, including ... IV and PD (assessing hand-object penetrations), PE ..."

    The update rule in Eq. (8) is a gradient descent on distance to the object surface plus MANO and normal-regularization terms, and the stated stopping rule declares success when there are no penetrations and contact regions are occupied. The IV, PD, and PE numbers reported in Table 1 are therefore, to a large extent, the direct optimization objective and stopping criterion rather than an independent predicted outcome. Reporting them as evidence of superiority is partially circular: the algorithm is constructed to minimize exactly these quantities, so low IV/PD/PE partly restates the definition of the process.

full rationale

The paper has no load-bearing self-citation chain; the main circularity is in the evaluation target and in reporting optimized objectives as predictions. The SHOWMe benchmark, which drives the headline SOTA claims, is not external: the authors state they refined the SHOWMe ground-truth hand meshes themselves, and the details are in a missing Appendix D. Since GravityDB is a post-processing deformable-hand optimizer, an author-produced deformation target can encode the same contact-attraction/MANO manifold the method is designed to reach. The GRAB and HO3D rows provide some independent external anchoring, but they are not enough to rescue the SHOWMe headline: Table 1 itself shows Ours is not state-of-the-art on MPVPE for GRAB (10.05 vs 9.23 for GeneOH) or HO3D (12.07 vs 11.31), so the paper's 'consistently outperforms' claim is internally overstated. Additionally, the SDE update and termination conditions explicitly minimize and then terminate on 'no penetrations' and occupied contact regions, so the IV/PD/PE improvements in Table 1 are largely by construction. Overall the core novelty—gravity-field SDE plus LLM-guided contact regions—has independent components, but its main quantitative validation on SHOWMe is self-referential, yielding a partial circularity score of 6.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The credit the paper earns is the integration of a Gaussian contact potential, SDE noise, MANO priors, and LLM contact masks. The cost is a set of hand-tuned coefficients and an evaluation target (refined SHOWMe GT) constructed by the authors. No new physical entities are introduced; the 'gravity field' is a mathematical potential, not a new force claim.

free parameters (8)
  • sigma_coarse = 0.2
    Receptive radius of the potential field in Stage 1; hand-tuned in Section 4.2 and controls how far attraction reaches.
  • sigma_fine = 0.05
    Receptive radius in Stage 2; hand-tuned in Section 4.2 to localize the attraction field.
  • lambda_1 (MANO weight) = 0.5
    Weight for the MANO shape prior in Eq. (5) and (7); tuned by cross-validation.
  • lambda_2 (normal alignment weight) = 0.2
    Weight for the contact-normal alignment loss in Eq. (6) and (7); tuned by cross-validation.
  • alpha (gravity step coefficient) = 0.01
    Euler-Maruyama step coefficient in Eq. (8); tuned by cross-validation.
  • epsilon and Tmax (stopping criteria) = 1e-4 and 150
    Gradient threshold and maximum iteration count; chosen by hand in Section 4.2.
  • Attraction strengths k_l and number of scales L
    Appear in Eq. (1) and (2) but are not given concrete values in the paper, despite controlling the gravity field behavior.
  • Noise schedule g(Xt)
    The SDE noise term in Eq. (3), (5), (7), (8) is described as spatially adaptive and annealed, but the schedule is not specified.
axioms (4)
  • ad hoc to paper The object surface can be modeled as a zero-potential manifold so attracting hand points toward it removes both interpenetration and gaps.
    This is the core modeling assumption behind Eq. (1)-(2). It maps physical contact to a geometric potential without modeling forces, friction, or object compliance.
  • domain assumption MANO hand model spans anatomically plausible hand shapes, so proximity to M(theta,beta) keeps deformations realistic.
    Used in LMANO in Eq. (4)-(5). If MANO lacks sufficient expressivity for the deformations the paper aims to capture, this constraint could suppress the very effect being studied.
  • ad hoc to paper The refined SHOWMe hand meshes are valid ground truth for realistic deformation.
    The authors state the SHOWMe ground truth was refined by them (Section 4.4). This is the benchmark assumption that makes the SOTA comparison meaningful.
  • domain assumption LLM-predicted contact masks from text and point clouds are reliable contact semantics.
    Required for Eq. (9) and the semantic weighting in Eq. (1). The LLM is fine-tuned on the authors' own annotations, so errors in those annotations propagate to the gravity field.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 12275 in / 11841 out tokens · 129898 ms · 2026-08-05T11:06:11.187766+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards Realistic Hand-Object Interaction with Gravity-Field Based Diffusion Bridge." pith.science (2026). https://pith.science/paper/SQCX4E3O

@misc{pith2026250903114,
  author       = {Pith},
  title        = {Pith review of: Towards Realistic Hand-Object Interaction with Gravity-Field Based Diffusion Bridge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQCX4E3O}},
  note         = {Machine review of arXiv:2509.03114}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing reconstruction or hand-object pose estimation methods are capable of producing coarse interaction states. However, due to the complex and diverse geometry of both human hands and objects, these approaches often suffer from interpenetration or leave noticeable gaps in regions that are supposed to be in contact. Moreover, the surface of a real human hand undergoes non-negligible deformations during interaction, which are difficult to capture and represent with previous methods. To tackle these challenges, we formulate hand-object interaction as an attraction-driven process and propose a Gravity-Field Based Diffusion Bridge (GravityDB) to simulate interactions between a deformable hand surface and rigid objects. Our approach effectively resolves the aforementioned issues by generating physically plausible interactions that are free of interpenetration, ensure stable grasping, and capture realistic hand deformations. Furthermore, we incorporate semantic information from textual descriptions to guide the construction of the gravitational field, enabling more semantically meaningful interaction regions. Extensive qualitative and quantitative experiments on multiple datasets demonstrate the effectiveness of our method.

Figures

Figures reproduced from arXiv: 2509.03114 by Jinlin Wu, Miao Xu, Xiangyu Zhu, Xusheng Liang, Zhen Lei, Zidu Wang.

Figure 1
Figure 1. Figure 1: We propose GravityDB to address three key challenges in hand-object interaction: (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method. (a) The LLM-Based Contact Regions Inference Module integrates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparative results with previous methods on SHOWMe. Since the reconstructed object [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization results of the LLM-Based Contact Regions Inference Module. Ours w/o Gravity employs only the contact region guidance, similar to the approach used in TOCH, without leveraging the gravity field to refine hand-object interactions. Ours w/o LMANO omit the anatomical regularization term λ1∇LMANOdt, removing constraints that ensure the deformation process remains within the man￾ifold of physically… 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

60 extracted references · 52 canonical work pages · 3 internal anchors

  1. [1]

    Generalizing from several related classifica- tion tasks to a new unlabeled sample

    Gilles Blanchard, Gyemin Lee, and Clayton Scott. Generalizing from several related classifica- tion tasks to a new unlabeled sample. Advances in neural information processing systems , 24, 2011

    work page 2011

  2. [2]

    Reconstructing hand-object interactions in the wild

    Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing hand-object interactions in the wild. In ICCV, pages 12417–12426, 2021

    work page 2021

  3. [3]

    gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction

    Zerui Chen, Shizhe Chen, Cordelia Schmid, and Ivan Laptev. gSDF: Geometry-driven signed distance functions for 3d hand-object reconstruction. In CVPR, pages 12890–12900, 2023

    work page 2023

  4. [4]

    Alignsdf: Pose-aligned signed distance fields for hand-object reconstruction

    Zerui Chen, Yana Hasson, Cordelia Schmid, and Ivan Laptev. Alignsdf: Pose-aligned signed distance fields for hand-object reconstruction. In European conference on computer vision , pages 231–248. Springer, 2022

    work page 2022

  5. [5]

    Improving diffusion models for inverse problems using manifold constraints

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. arXiv preprint arXiv:2206.00941, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    GanHand: Predicting human grasp affordances in multi-object scenes

    Enric Corona, Albert Pumarola, Guillem Alenyà, Francesc Moreno-Noguer, and Grégory Rogez. GanHand: Predicting human grasp affordances in multi-object scenes. In CVPR, pages 5030–5040, 2020

    work page 2020

  7. [7]

    Domain gener- alization via model-agnostic learning of semantic features

    Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain gener- alization via model-agnostic learning of semantic features. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  8. [8]

    Hold: Category-agnostic 3d reconstruction of interacting hands and objects from video

    Zicong Fan, Maria Parelli, Maria Eleni Kadoglou, Xu Chen, Muhammed Kocabas, Michael J Black, and Otmar Hilliges. 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, pages 494–504, 2024

    work page 2024

  9. [9]

    Black, and Otmar Hilliges

    Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  10. [10]

    Imos: Intent-driven full-body motion synthesis for human-object interactions

    Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Computer Graphics F orum, volume 42, pages 1–12. Wiley Online Library, 2023

    work page 2023

  11. [11]

    Contactopt: Optimizing contact to improve grasps

    Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh V o, Samarth Brahmbhatt, and Charles C Kemp. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1471–1481, 2021

    work page 2021

  12. [12]

    Interaction Replica: Tracking Human-Object Interaction and Scene Changes From Human Motion

    Vladimir Guzov, Torsten Sattler, and Gerard Pons-Moll. Visually plausible human-object interaction capture from wearable sensors. arXiv preprint arXiv:2205.02830, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  13. [13]

    Honnotate: A method for 3d annotation of hand and object poses

    Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 10

    work page 2020

  14. [14]

    Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation

    Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090–11100, 2022

    work page 2022

  15. [15]

    Leveraging photometric consistency over time for sparsely supervised hand-object reconstruc- tion

    Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruc- tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 571–580, 2020

    work page 2020

  16. [16]

    Leveraging photometric consistency over time for sparsely supervised hand-object reconstruc- tion

    Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruc- tion. In CVPR, pages 568–577, 2020

    work page 2020

  17. [17]

    Towards unconstrained joint hand-object reconstruction from rgb videos

    Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. pages 659–668. IEEE, 2021

    work page 2021

  18. [18]

    Learning joint reconstruction of hands and manipulated objects

    Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11807–11816, 2019

    work page 2019

  19. [19]

    Black, Ivan Laptev, and Cordelia Schmid

    Yana Hasson, Gül Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, pages 11807–11816, 2019

    work page 2019

  20. [20]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    2022

  21. [21]

    Reconstructing hand-held objects from monocular video

    Di Huang, Xiaopeng Ji, Xingyi He, Jiaming Sun, Tong He, Qing Shuai, Wanli Ouyang, and Xiaowei Zhou. Reconstructing hand-held objects from monocular video. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022

    work page 2022

  22. [22]

    Transferability in deep learning: A survey

    Junguang Jiang, Yang Shu, Jianmin Wang, and Mingsheng Long. Transferability in deep learning: A survey. arXiv preprint arXiv:2201.05867, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    Black, Krikamol Muandet, and Siyu Tang

    Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping Field: Learning implicit representations for human grasps. pages 333–344, 2020

    work page 2020

  24. [24]

    H2o: Two hands manipulating objects for first person interaction recognition

    Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138–10148, 2021

    work page 2021

  25. [25]

    Task-oriented human-object interactions generation with implicit neural representations

    Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. arXiv preprint arXiv:2303.13129, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Semi-supervised 3D hand-object poses estimation with interactions in time

    Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3D hand-object poses estimation with interactions in time. In CVPR, pages 14687–14697, 2021

    work page 2021

  27. [27]

    Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion

    Xueyi Liu and Li Yi. Geneoh diffusion: Towards generalizable hand-object interaction denoising via denoising diffusion. arXiv preprint arXiv:2402.14810, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    A variational perspective on solving inverse problems with diffusion models

    Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations , 2021

    work page 2021

  30. [30]

    Domain generalization via invariant feature representation

    Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International conference on machine learning , pages 10–18. PMLR, 2013. 11

    work page 2013

  31. [31]

    Comparison of Spatio-Temporal Hand Pose Denoising Models

    Johnny Núñez. Comparison of Spatio-Temporal Hand Pose Denoising Models . PhD thesis, Universitat DE Barcelona, 2022

    work page 2022

  32. [32]

    3D Reconstruction of Objects in Hands without Real World 3D Supervision

    Aditya Prakash, Matthew Chang, Matthew Jin, and Saurabh Gupta. Learning hand-held object reconstruction from in-the-wild videos. arXiv, 2305.03036, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  33. [33]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017

    work page 2017

  34. [34]

    Novel-view Synthesis and Pose Estimation for Hand-Object Interaction from Sparse Views

    Wentian Qu, Zhaopeng Cui, Yinda Zhang, Chenyu Meng, Cuixia Ma, Xiaoming Deng, and Hongan Wang. Novel-view synthesis and pose estimation for hand-object interaction from sparse views. arXiv preprint arXiv:2308.11198, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  35. [35]

    Batch normalization embeddings for deep domain generalization

    Mattia Segu, Alessio Tonioni, and Federico Tombari. Batch normalization embeddings for deep domain generalization. Pattern Recognition, 135:109115, 2023

    work page 2023

  36. [36]

    Domain adversarial neural networks for domain generalization: When it works and how to improve

    Anthony Sicilia, Xingchen Zhao, and Seong Jae Hwang. Domain adversarial neural networks for domain generalization: When it works and how to improve. Machine Learning, pages 1–37, 2023

    work page 2023

  37. [37]

    Solv- ing inverse problems with latent diffusion models via hard data consistency

    Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solv- ing inverse problems with latent diffusion models via hard data consistency. arXiv preprint arXiv:2307.08123, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction

    Anilkumar Swamy, Vincent Leroy, Philippe Weinzaepfel, Fabien Baradel, Salma Galaaoui, Romain Brégier, Matthieu Armando, Jean-Sebastien Franco, and Grégory Rogez. SHOWMe: Benchmarking object-agnostic hand-object 3d reconstruction. In ICCV, pages 1935–1944, 2023

    work page 1935

  39. [39]

    Grab: A dataset of whole-body human grasping of objects

    Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. 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 , pages 581–600. Springer, 2020

    work page 2020

  40. [40]

    H+O: Unified egocentric recognition of 3D hand-object poses and interactions

    Bugra Tekin, Federica Bogo, and Marc Pollefeys. H+O: Unified egocentric recognition of 3D hand-object poses and interactions. In CVPR, pages 4511–4520, 2019

    work page 2019

  41. [41]

    Flex: Full-body grasping without full- body grasps

    Purva Tendulkar, Dídac Surís, and Carl V ondrick. Flex: Full-body grasping without full- body grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21179–21189, 2023

    work page 2023

  42. [42]

    Pose-ndf: Modeling human pose manifolds with neural distance fields

    Garvita Tiwari, Dimitrije Anti´c, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In Com- puter Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part V, pages 572–589. Springer, 2022

    work page 2022

  43. [43]

    Collaborative learning for hand and object reconstruction with attention-guided graph convolution

    Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, pages 1664–1674, 2022

    work page 2022

  44. [44]

    Plug-and-play diffusion features for text-driven image-to-image translation

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921–1930, 2023

    work page 1921

  45. [45]

    Generalizing to unseen domains: A survey on domain generaliza- tion

    Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generaliza- tion. IEEE Transactions on Knowledge and Data Engineering , 2022

    work page 2022

  46. [46]

    Sharpness-aware gradient matching for domain generalization

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769–3778, 2023. 12

    work page 2023

  47. [47]

    Saga: Stochastic whole-body grasping with contact

    Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI , pages 257–274. Springer, 2022

    work page 2022

  48. [48]

    Chore: Contact, human and object reconstruction from a single rgb image

    Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II , pages 125–145. Springer, 2022

    work page 2022

  49. [49]

    CPF: Learning a contact potential field to model the hand-object interaction

    Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. CPF: Learning a contact potential field to model the hand-object interaction. In ICCV, pages 11097–11106, 2021

    work page 2021

  50. [50]

    CPF: Learning a contact potential field to model the hand-object interaction

    Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. CPF: Learning a contact potential field to model the hand-object interaction. In ICCV, 2021

    work page 2021

  51. [51]

    What’s in your hands? 3d reconstruction of generic objects in hands

    Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What’s in your hands? 3d reconstruction of generic objects in hands. In CVPR, 2022

    work page 2022

  52. [52]

    What’s in your hands? 3d reconstruction of generic objects in hands

    Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. 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, pages 3895–3905, 2022

    work page 2022

  53. [53]

    Diffusion-guided reconstruc- tion of everyday hand-object interaction clips

    Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruc- tion of everyday hand-object interaction clips. In ICCV, 2023

    work page 2023

  54. [54]

    Diffusion-guided reconstruc- tion of everyday hand-object interaction clips

    Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruc- tion of everyday hand-object interaction clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19717–19728, 2023

    work page 2023

  55. [55]

    Feder- ated domain generalization with generalization adjustment

    Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Feder- ated domain generalization with generalization adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3954–3963, 2023

    work page 2023

  56. [56]

    The wanderings of odysseus in 3d scenes

    Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481–20491, 2022

    work page 2022

  57. [57]

    Domain generalization with mixstyle

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021

    Pith/arXiv arXiv 2021

  58. [58]

    Stgae: Spatial-temporal graph auto-encoder for hand motion denoising

    Kanglei Zhou, Zhiyuan Cheng, Hubert PH Shum, Frederick WB Li, and Xiaohui Liang. Stgae: Spatial-temporal graph auto-encoder for hand motion denoising. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR) , pages 41–49. IEEE, 2021

    work page 2021

  59. [59]

    Toch: Spatio- temporal object correspondence to hand for motion refinement

    Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio- temporal object correspondence to hand for motion refinement. In European Conference on Computer Vision (ECCV). Springer, October 2022

    work page 2022

  60. [60]

    Monocular real-time hand shape and motion capture using multi-modal data

    Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In CVPR, pages 5345–5354, 2020. 13

    work page 2020