Pith. sign in

REVIEW 3 major objections 6 minor 50 references

OASIS reconstructs an animatable 3D hand avatar from a single image using 3D Gaussian Splatting, beating NeRF-based one-shot methods in both fidelity and speed.

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

2026-08-03 03:02 UTC pith:4NZXK4PL

load-bearing objection A credible engineering advance that is oversold on occlusion: the core visibility-conditioned mechanism is only evaluated on a no-occlusion sequence, with a small ablation delta. the 3 major comments →

arxiv 2607.29633 v1 pith:4NZXK4PL submitted 2026-07-31 cs.CV

OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting

Zhisheng Han , Shiyao Wu , Jiayan Qiu , Yakun Ju , Lu Liu , Le Zhang , Pengfei Feng , Huiyu Zhou
show 1 more author
Zheheng Jiang
This is my paper
classification cs.CV
keywords hand avatar reconstructionsingle-image reconstruction3D Gaussian Splattingone-shot adaptationocclusion-aware attentionvisibility conditioningFeature-on-Mesharticulated hand deformation
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 claims that single-image hand avatar reconstruction, long considered ill-posed because of severe self-occlusion and complex articulation, can be solved accurately and efficiently with an explicit 3D Gaussian Splatting pipeline rather than implicit volumetric rendering. To make sparse image evidence usable, OASIS aligns image features with 3D hand geometry and compresses them into visual evidence tokens; a visibility-conditioned attention mechanism then routes each 3D point to local image cues when it is visible and to global context when it is occluded. A mesh-face feature representation ties Gaussian deformation to local surface stretching, and a one-shot adaptation stage -- global color calibration followed by low-rank fine-tuning -- injects the target identity while preserving learned priors for unseen regions. If the reported results hold, the method delivers photorealistic, re-animatable hand avatars from a single photo in about five minutes with real-time rendering, a practical improvement over NeRF-based one-shot methods that take an hour and render at well below one frame per second.

Core claim

OASIS is a 3D Gaussian Splatting framework for single-image hand avatar reconstruction. It projects posed 3D hand points onto the image and compresses the aligned features into visual evidence tokens; a z-buffer visibility score then biases point-to-image cross-attention so visible points use local tokens and occluded points use a global token. The resulting features are lifted onto mesh faces, and Gaussians follow face-level deformation to capture non-rigid articulation. A one-shot stage -- color calibration plus low-rank fine-tuning -- personalizes the prior. The paper reports that this outperforms existing one-shot methods on InterHand2.6M across PSNR, SSIM, and LPIPS.

What carries the argument

Three modules carry the argument. Geometry-aligned visual evidence tokens project posed 3D points into the image and compress local features via learnable queries. A z-buffer-based soft visibility score biases cross-attention logits, routing visible points to local tokens and occluded points to a global token. Feature-on-Mesh interpolates vertex features onto mesh faces with barycentric coordinates, so decoded Gaussians follow face-level rotation and scaling. One-shot adaptation adds color calibration and low-rank fine-tuning of the Gaussian decoder.

Load-bearing premise

The load-bearing premise is that the off-the-shelf hand pose and camera estimator produces MANO pose, shape, and intrinsics that align the image with the 3D hand accurately enough; if the estimated geometry is wrong, the projected evidence, visibility scores, and one-shot fitting all start from misaligned geometry and the reconstruction degrades.

What would settle it

Run OASIS on InterHand2.6M test sequences using ground-truth MANO pose and camera instead of the off-the-shelf estimator, and compare PSNR/SSIM/LPIPS with the reported numbers: a substantial drop would confirm the dependency on the estimator, while a negligible drop would weaken the claimed limitation.

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

If this is right

  • If the reported results hold, single-image hand avatars become deployable: adaptation takes about five minutes and rendering runs at roughly 390 FPS, compared with about one hour and 0.16 FPS for the NeRF-based one-shot baseline.
  • High-frequency appearance -- fingernails, palm wrinkles, rings -- is preserved under challenging poses and in-the-wild images, not just in controlled benchmark conditions.
  • The same prior-plus-personalization pipeline supports text-to-avatar generation and texture editing from a single image, extending beyond plain reconstruction.
  • The no-finetuning variant already surpasses the one-shot baselines it is compared with, indicating that the learned prior alone carries much of the reconstruction quality.
  • The method is robust across unseen poses and viewpoints, with smoother transitions between observed and unobserved regions than the NeRF-based baseline.

Where Pith is reading between the lines

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

  • Editorial inference: the z-buffer visibility prior is a general remedy for any articulated, self-occluding object reconstructed from a single view; it could be transferred to full-body or hand-object interaction avatars.
  • Editorial inference: the Feature-on-Mesh scheme suggests that vertex-based 3DGS avatar systems often produce holes under stretching because they ignore face-level deformation; face-level carriers may fix this artifact broadly.
  • Editorial inference: because the pipeline relies on an off-the-shelf hand pose and camera estimator, feeding ground-truth MANO parameters on the test set would quantify how much of the reported accuracy is inherited from the estimator rather than from OASIS itself.
  • Editorial inference: the paper bakes illumination into texture; combining OASIS with intrinsic decomposition would make the avatars re-lighting-capable, a natural next step the authors note as future work.

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 / 6 minor

Summary. The paper proposes OASIS, a 3D Gaussian Splatting framework for reconstructing an animatable hand avatar from a single RGB image. The method consists of three main components: geometry-aligned visual evidence tokens (VETs) that project image features onto the MANO-HD hand mesh and compress them into compact tokens; a visibility-conditioned point-image attention (VPIA) that modulates point-to-image attention using a z-buffer soft-visibility score, directing visible points to local tokens and occluded points to a global token; and a Feature-on-Mesh (FoM) representation that stores features on mesh faces and deforms Gaussians according to local mesh stretching. The model is pretrained on multiple identities from InterHand2.6M and then adapted to a target image via color calibration and LoRA finetuning. Experiments compare against NeRF-based one-shot and video-based baselines on InterHand2.6M and HanCo, reporting higher PSNR/SSIM/LPIPS and much faster rendering and adaptation. Qualitative results are shown for in-the-wild images and downstream text-to-avatar/texture-editing applications.

Significance. If the claims hold, OASIS is a strong contribution: it is the first 3DGS-based one-shot hand avatar method with real-time rendering (~390 FPS) and a practical ~5-minute personalization, in contrast to the ~1-hour optimization and ~0.16 FPS rendering of the NeRF-based OHTA. The geometry-aligned tokenization and FoM representation are sensible and potentially reusable for other articulated single-view reconstruction tasks. The authors are transparent about testing-time fitting and acknowledge upstream estimator dependency in the limitations. The main weakness is that the paper's central novel component, occlusion-aware attention, is not quantitatively evaluated under occlusion, and several ablations show small single-run differences without statistical support. The empirical record therefore does not yet firmly establish the occlusion-awareness claims, although the overall system shows promising results.

major comments (3)
  1. [§4.2 / Appendix C.2] The main quantitative evaluation is performed on sequence 'test/Capture0/ROM03_RT_No_Occlusion' (App. C.2), which is explicitly a no-occlusion sequence. The title and abstract emphasize 'occlusion-aware' reconstruction, and VPIA is motivated as necessary for 'severe self-occlusion' (§3.3). Yet no quantitative evaluation is provided on occluded sequences or on an occlusion-stratified split. The only occlusion-related ablation (Table 2, rows a vs c) shows a 0.28 PSNR gain and identical SSIM (0.956). This does not demonstrate that the visibility-conditioned mechanism provides the claimed benefit under occlusion. Please evaluate on occluded InterHand2.6M sequences (e.g., ROM occlusion sequences) or provide a visibility/occlusion-based breakdown of metrics.
  2. [Tables 1–4] All quantitative tables report single-run metrics without error bars or significance tests. Several ablation differences are very small relative to the likely run-to-run variance: Table 2 full vs w/o VPIA differs by 0.28 PSNR and 0.64 LPIPS with SSIM unchanged; Table 2 c.1 vs c differs by 0.03 PSNR; Table 4 d vs a differs by 0.36 PSNR. The text claims these components are 'crucial' (§4.3), but such small single-run deltas are not sufficient to support that claim. Please report mean±std over multiple subjects/views, or provide per-sequence results with statistical significance tests, especially for the ablation tables.
  3. [§4.4 / Appendix C.5] The claim of superior performance 'in-the-wild scenarios' is supported only by qualitative comparisons (Fig. 6 and Fig. 7). Appendix C.5, cited as 'quantitative comparisons for robustness,' actually reports per-image PSNR/LPIPS on InterHand2.6M, not on in-the-wild data. No quantitative evaluation on COCO-Hand or WHIM is provided. Since in-the-wild generalization is a stated headline advantage, the authors should either add quantitative results on these datasets (with pseudo-ground-truth or a user study) or moderate the claim to 'qualitative robustness on in-the-wild images.'
minor comments (6)
  1. [App. D(a)] The dependency on off-the-shelf WiLoR estimates is acknowledged and is a reasonable limitation. Consider quantifying sensitivity by perturbing pose/camera estimates or reporting performance on a subset with high WiLoR error.
  2. [Eq. (20)] There is a typesetting error: 'smax_i, smax_i' should presumably be 'smax_i, smin_i' (max and min scaling values).
  3. [App. C.6] The text refers to 'tab. 11 #c' when discussing color-calibration ablation; this should be Table 4.
  4. [Table 2 / Table 4] SSIM is essentially saturated at 0.955–0.956 across several ablations; consider also reporting a metric that is more sensitive to fine texture, e.g., PSNR on high-frequency regions, to better differentiate methods.
  5. [§4.2] Minor language issue: 'our one-shot results are the most competitive with them' is ambiguous; consider 'our one-shot results are competitive with these video-supervised methods.'
  6. [Author affiliations] Typo: 'Birminghan' should be 'Birmingham.'

Circularity Check

0 steps flagged

No significant circularity: the one-shot result is explicitly an optimization/fitting stage, the prior net is trained on disjoint subjects, and all load-bearing citations are external.

full rationale

This is an empirical systems paper, not a derivation with a hidden definitional loop. The one-shot pipeline is explicitly presented as test-time personalization: color calibration (Eq. 22) and LoRA finetuning are optimized against the input image after the prior net is trained on 21 InterHand2.6M training subjects, and evaluation is performed on InterHand2.6M test sequence 'test/Capture0/ROM03_RT_No_Occlusion' (App. C.2), so the target subject is not in the training set. The no-finetuning variant 'Ours§' is explicitly labeled 'testing without finetuning,' so the reported numbers are not a fitted parameter renamed as a prediction. The visibility score in Eq. 4 is computed from the z-buffer of the posed MANO-HD geometry and used as an input bias to attention; it is not derived from the output reconstruction and no equation in the paper reduces to its own input. The Feature-on-Mesh representation (Eq. 7) is a learnable barycentric interpolation of vertex features, not a repackaging of the evaluation metric. Load-bearing components such as MANO, DINOv2, WiLoR, 3DGS, InterHand2.6M, and the OHTA protocol are all external prior work; there is no same-author self-citation chain being used to justify a key claim. The skeptic's concern that the occlusion-aware claim is mainly tested on a no-occlusion sequence and that in-the-wild evidence is qualitative is a valid evidentiary gap about whether the headline occlusion claim is fully supported, but it is not circularity: the reported gains do not reduce, by construction, to the inputs of the method. Therefore the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 4 invented entities

The central method is a learned pipeline: it takes MANO geometry and DINOv2 features as inputs, and learns to map them to Gaussians. The free parameters are mostly training hyperparameters; the load-bearing assumptions are that the parametric hand model, the off-the-shelf pose/camera estimator, the visual backbone, and the training distribution are all good enough for arbitrary inputs. The invented entities are internal neural representation designs, not physical postulates, and they carry no independent evidence outside the paper's own ablations.

free parameters (6)
  • visibility temperature τ (Eq. 4)
    Controls the softness of the sigmoid mapping from z-buffer depth residual to visibility score; directly shapes the VPIA bias, the paper's core novelty. No value is reported.
  • visibility bias strengths β_local, β_global (Eq. 5)
    Sets the magnitude of the local/global attention bias in VPIA; not reported, though ablation c.1 shows the bias matters.
  • Pretraining loss weights λ1..λ4, λp, λs (Eq. 14, Eq. 21) = 10, 1, 1, 5, 1, 0.5
    Hand-chosen weights for reconstruction and regularization losses in Sec 4.1.
  • One-shot color-calibration weights λμ, λσ (Eq. 22) = 20, 1
    Hand-chosen weights for mean/std color statistics losses during one-shot adaptation.
  • Regularization thresholds d and r (Eq. 19-20)
    Empirically determined thresholds to keep Gaussians near the MANO surface and avoid thin primitives; values are not reported.
  • Number of tokens N_vet and points N_points
    Architectural capacities for the tokenization queries and Gaussian count, chosen by hand; not reported, affecting fidelity-vs-efficiency tradeoff.
axioms (5)
  • domain assumption MANO-HD parametric hand model is an accurate topological and LBS deformation prior for all target hands.
    All 3D points, mesh faces, and Gaussian initialization live on MANO-HD; hands that deviate from this parametric family cannot be captured. Invoked in Sec 3.2 and Sec A.
  • domain assumption The off-the-shelf WiLoR estimator provides sufficiently accurate MANO pose, shape, camera, and handedness.
    The geometry-aligned projection (Sec 3.2.1), z-buffer visibility test (Eq. 4), and one-shot fitting all start from these estimates; the paper's own limitation section D(a) confirms failure under poor estimates.
  • domain assumption DINOv2 features contain enough appearance information for hand texture transfer.
    VETs are built solely from DINOv2 features sampled around projected points; if the backbone misses hand-specific high-frequency cues, the VETs cannot recover them. Sec 3.2.
  • domain assumption The 21 InterHand2.6M training subjects form a prior that generalizes to unseen identities and in-the-wild inputs.
    The prior network is trained on these subjects; one-shot adaptation assumes this initialization is useful for any target. Sec 4.1.
  • standard math Linear blend skinning and differentiable 3D Gaussian splatting correctly model hand pose deformation and image formation.
    The avatar is animated and rendered via LBS and the 3DGS rasterizer (Eq. 10-12); any modeling error in these components propagates directly to reconstruction quality.
invented entities (4)
  • Geometry-aligned Visual Evidence Tokens (VETs) no independent evidence
    purpose: Compact tokens summarizing point-aligned DINOv2 features via learnable cross-attention; the primary image-evidence carrier.
    Introduced in Sec 3.2; validated only by in-paper ablation (Table 2 #b), no external falsifiable handle.
  • Visibility-conditioned Point-Image Attention (VPIA) bias no independent evidence
    purpose: Logit-level additive bias modulated by z-buffer visibility to route local vs global evidence.
    Introduced in Sec 3.3; ablation c.1 supports it within the paper, but there is no independent benchmark of the bias mechanism.
  • Feature-on-Mesh (FoM) representation no independent evidence
    purpose: Face-level barycentric feature carrier that ties Gaussian deformation to mesh-face stretching.
    Introduced in Sec 3.4; variant of mesh-embedded Gaussian ideas (e.g., SplattingAvatar), validated only by ablation #d.
  • Occlusion-aware Gaussian Features (OGF) no independent evidence
    purpose: The fused geometric+visual feature vectors used by the Gaussian decoder.
    Output of the VPIA module (Sec 3.3); an internal representation with no external validity check.

pith-pipeline@v1.3.0-daily-deepseek · 18619 in / 14834 out tokens · 122439 ms · 2026-08-03T03:02:17.146307+00:00 · methodology

0 comments
read the original abstract

Single-image 3D hand avatar reconstruction is fundamentally ill-posed and particularly challenging due to limited visual evidence under severe self-occlusion and the complex pose-dependent deformation of highly articulated hands. Existing methods predominantly rely on implicit NeRF-style representations, whose volumetric fitting is computationally expensive and often struggles to preserve fine-grained hand details. In this work, we present OASIS, a tailored 3D Gaussian Splatting framework for single-image hand avatar reconstruction. To faithfully encode sparse image-specific appearance cues in single-view reconstruction, we construct geometry-aligned visual evidence tokens by explicitly aligning input image observations with 3D hand geometry and context-adaptively tokenizing the resulting visual evidence. Since severe self-occlusion makes the reliability of image evidence inherently visibility-dependent, we introduce a visibility-conditioned point-image attention to reliably transfer visual evidence to geometric tokens, yielding occlusion-aware Gaussian features for faithful and robust reconstruction. To further capture non-rigid deformation of articulated hands, we introduce a Feature-on-Mesh representation to enable Gaussian deformation to be guided by local surface stretching. Under this framework, we adopt a one-shot adaptation scheme that learns a shared hand prior from multi-identity training data and then fits it to a target image for target-specific reconstruction. Extensive experiments show that OASIS outperforms existing baselines in both visual fidelity and efficiency across challenging poses and in-the-wild scenarios, and further demonstrates strong versatility in downstream applications such as text-to-avatar generation and texture editing.

Figures

Figures reproduced from arXiv: 2607.29633 by Huiyu Zhou, Jiayan Qiu, Le Zhang, Lu Liu, Pengfei Feng, Shiyao Wu, Yakun Ju, Zheheng Jiang, Zhisheng Han.

Figure 1
Figure 1. Figure 1: We introduce OASIS, a 3DGS-based one-shot hand avatar reconstruction model with efficient adaptation (∼5 minutes) and real-time rendering (∼390 FPS). We showcase a gallery of one-shot hand avatars that highlights both robustness and diversity: it reconstructs high-fidelity hands under challenging poses, complex textures, and extreme side-view observations, while also supporting human-image inputs and downs… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed OASIS. Given a single input image, we construct geometry-aligned VETs to encode image-specific appearance [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the proposed Visibility-conditioned Point [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with state-of-the-art methods on In [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual results for ablation studies of OASIS on InterHand2.6M dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: In the wild visualization from the HanCo [42] dataset, COCO-Hand [20] dataset, and WHIM [23] dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison with [41] on text-to-avatar application. The first column is the input reference images, and the rendered target views are displayed on the right side, respectively. The following two visual results are obtained by animating the reconstructed 3D hand avatar to different poses [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on HanCo [42]. The white box indicates the input view. Method PSNR↑ LPIPS*↓ SSIM↑ HandAvatar [3] 24.68 8.25 0.896 OHTA [41] 25.24 7.18 0.923 Ours 25.86 6.76 0.974 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Points visibility and VPIA attention visualization. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Robustness towards different input images. The images [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual results for ablation study with in-the-wild input. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Failure cases. visible regions and shifts to global context for occluded regions, which is consistent with the design of VPIA. One-shot Reconstruction. To validate each component’s effec￾tiveness, we perform ablation studies under the one-shot setting. The quantitative results are summarized in tab. 4 on the Inter￾Hand2.6M dataset, and the qualitative results are performed on in-the-wild images, displayed… 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

50 extracted references · 3 linked inside Pith

  1. [1]

    Jinnan Chen, Chen Li, Jianfeng Zhang, Lingting Zhu, Buzhen Huang, Hanlin Chen, and Gim Hee Lee. 2024. Generalizable Human Gaussians from Single-View Images.arXiv preprint arXiv:2406.06050(2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Wenyue Chen, Peng Li, Wangguandong Zheng, Chengfeng Zhao, Mengfei Li, Yaolong Zhu, Zhiyang Dou, Ronggang Wang, and Yuan Liu. 2025. SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Recon- struction.arXiv preprint arXiv:2510.07723(2025)

    arXiv 2025

  3. [3]

    Xingyu Chen, Baoyuan Wang, and Heung-Yeung Shum. 2023. HandAvatar: Free- Pose Hand Animation and Rendering from Monocular Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8683–8693

    2023

  4. [4]

    Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. 2022. Lisa: Learning implicit shape and appearance of hands. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20533–20543

    2022

  5. [5]

    Haoye Dong, Aviral Chharia, Wenbo Gou, Francisco Vicente Carrasco, and Fer- nando D De la Torre. 2024. Hamba: Single-view 3D Hand Reconstruction with Graph-guided Bi-Scanning Mamba.Advances in Neural Information Processing Systems37 (2024), 2127–2160

    2024

  6. [6]

    Zhiyang Guo, Wengang Zhou, Min Wang, Li Li, and Houqiang Li. 2023. Hand- NeRF: Neural Radiance Fields for Animatable Interacting Hands. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 21078–21087

    2023

  7. [7]

    Yisheng He, Xiaodong Gu, Xiaodan Ye, Chao Xu, Zhengyi Zhao, Yuan Dong, Weihao Yuan, Zilong Dong, and Liefeng Bo. 2025. LAM: Large Avatar Model for One-shot Animatable Gaussian Head. InProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. 1–13

    2025

  8. [8]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.Iclr1, 2 (2022), 3

    2022

  9. [9]

    Shoukang Hu, Tao Hu, and Ziwei Liu. 2024. Gauhuman: Articulated gaussian splatting from monocular human videos. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20418–20431

    2024

  10. [10]

    Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. 2022. SelfRecon: Self Reconstruction of Your Digital Avatar from Monocular Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5605–5615

    2022

  11. [11]

    Korrawe Kar., Sergey Prokudin, Otmar Hilliges, and Siyu Tang. 2023. HARP: Personalized Hand Reconstruction from a Monocular RGB Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12802– 12813

    2023

  12. [12]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al

  13. [13]

    Minje Kim and Tae-Kyun Kim. 2024. Bitt: Bi-directional texture reconstruction of interacting two hands from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10726–10735

    2024

  14. [14]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

    Pith/arXiv arXiv 2014

  15. [15]

    Yuwei Li, Longwen Zhang, Zesong Qiu, Yingwenqi Jiang, Nianyi Li, Yuexin Ma, Yuyao Zhang, Lan Xu, and Jingyi Yu. 2022. Nimble: a non-rigid hand model with bones and muscles.ACM Transactions on Graphics (TOG)41, 4 (2022), 1–16

    2022

  16. [16]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. NeRF: Representing scenes as neural radiance fields for view synthesis.Commun. ACM65, 1 (2021), 99–106

    2021

  17. [17]

    Gyeongsik Moon, Weipeng Xu, Rohan Joshi, Chenglei Wu, and Takaaki Shiratori

  18. [18]

    Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee

  19. [19]

    Akshay Mundra, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mo- hamed Elgharib. 2023. LiveHand: Real-Time and Photorealistic Neural Hand Rendering. InProceedings of the IEEE/CVF International Conference on Computer Vision. 18035–18045

    2023

  20. [20]

    Supreeth Narasimhaswamy, Zhengwei Wei, Yang Wang, Justin Zhang, and Minh Hoai. 2019. Contextual attention for hand detection in the wild. InProceedings of the IEEE/CVF international conference on computer vision. 9567–9576

    2019

  21. [21]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2024. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research Journal(2024)

    2024

  22. [22]

    Rolandos Alexandros Potamias, Stylianos Ploumpis, Stylianos Moschoglou, Vasileios Triantafyllou, and Stefanos Zafeiriou. 2023. Handy: Towards a High- Fidelity 3D Hand Shape and Appearance Model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4670–4680

    2023

  23. [23]

    Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. 2025. Wilor: End-to-end 3d hand localization and reconstruction in- the-wild. InProceedings of the Computer Vision and Pattern Recognition Conference. 12242–12254

    2025

  24. [24]

    Lorenza Prospero, Abdullah Hamdi, Joao F Henriques, and Christian Rupprecht

  25. [25]

    Neng Qian, Jiayi Wang, Franziska Mueller, Florian Bernard, Vladislav Golyanik, and Christian Theobalt. 2020. Html: A parametric hand texture model for 3d hand reconstruction and personalization. InEuropean Conference on Computer Vision. Springer, 54–71

    2020

  26. [26]

    Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang

  27. [27]

    Lingteng Qiu, Xiaodong Gu, Peihao Li, Qi Zuo, Weichao Shen, Junfei Zhang, Kejie Qiu, Weihao Yuan, Guanying Chen, Zilong Dong, and Liefeng Bo. 2025. LHM: Large Animatable Human Reconstruction Model for Single Image to 3D in Seconds. InProceedings of the IEEE/CVF International Conference on Computer Vision. 14184–14194

    2025

  28. [28]

    Lingteng Qiu, Shenhao Zhu, Qi Zuo, Xiaodong Gu, Yuan Dong, Junfei Zhang, Chao Xu, Zhe Li, Weihao Yuan, Liefeng Bo, et al . 2025. Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction. InProceedings of the Computer Vision and Pattern Recognition Conference. 21148– 21158

    2025

  29. [29]

    Javier Romero, Dimitrios Tzionas, and Michael J Black. 2017. Embodied hands: modeling and capturing hands and bodies together.ACM Transactions on Graphics (TOG)36, 6 (2017), 1–17

    2017

  30. [30]

    Umme Sara, Morium Akter, Mohammad Shorif Uddin, et al. 2019. Image quality assessment through FSIM, SSIM, MSE and PSNR—a comparative study.Journal of Computer and Communications7, 3 (2019), 8–18

    2019

  31. [31]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5020–5030

  32. [32]

    Zhigang Tu, Zhisheng Huang, Yujin Chen, Di Kang, Linchao Bao, Bisheng Yang, and Junsong Yuan. 2023. Consistent 3d hand reconstruction in video via self- supervised learning.IEEE Transactions on Pattern Analysis and Machine Intelli- gence45, 8 (2023), 9469–9485

    2023

  33. [33]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 4 (2004), 600–612

    2004

  34. [34]

    Srinivasan, Jonathan T

    Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. 2022. HumanNeRF: Free-Viewpoint Rendering of Moving People from Monocular Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16210–16220

    2022

  35. [35]

    Yufan Wu, Xuanhong Chen, Wen Li, Shunran Jia, Hualiang Wei, Kairui Feng, Jialiang Chen, Yuhan Li, Ang He, Weimin Zhang, et al. 2025. SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors. InProceedings of the Computer Vision and Pattern Recognition Conference. 5571–5580

    2025

  36. [36]

    Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. 2024. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1606–1616

    2024

  37. [37]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2024. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072(2024)

    Pith/arXiv arXiv 2024

  38. [38]

    Dongbin Zhang, Yunfei Liu, Lijian Lin, Ye Zhu, Yang Li, Minghan Qin, Yu Li, and Haoqian Wang. 2025. GUAVA: Generalizable Upper Body 3D Gaussian Avatar. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 14205– 14217

    2025

  39. [39]

    Ruiqi Zhang and Jie Chen. 2024. Mesh-centric gaussian splatting for human avatar modelling with real-time dynamic mesh reconstruction. InProceedings of the 32nd ACM International Conference on Multimedia. 6823–6832

    2024

  40. [40]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  41. [41]

    Yue Wu, Yufan Wu, Wen Li, Yuxi Lu, Kairui Feng, and Xuanhong Chen. 2025. FastAvatar: Towards Unified Fast High-Fidelity 3D Avatar Reconstruction with Large Gaussian Reconstruction Transformers.arXiv preprint arXiv:2508.19754 (2025)

    arXiv 2025

  42. [42]

    Christian Zimmermann, Max Argus, and Thomas Brox. 2021. Contrastive repre- sentation learning for hand shape estimation. InDAGM German Conference on Pattern Recognition. Springer, 250–264. OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Appendix A Preliminary 3D ...

    2021

  43. [47]

    Xiaozheng Zheng, Chao Wen, Zhuo Su, Zeran Xu, Zhaohu Li, Yang Zhao, and Zhou Xue. 2024. OHTA: One-Shot Hand Avatar via Data-Driven Implicit Priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 799–810

    2024

  44. [49]

    to use the subject ‘train/Capture0’, ‘train/Capture1’, ‘train/Capture2’, ‘train/Capture3’, ‘train/Capture5’, ‘train/Capture6’, ‘train/Capture7’, ‘train/Capture8’, ‘train/Capture9’, ‘train/Capture10’, ‘train/Capture11’, ‘train/Capture12’, ‘train/Capture13’, ‘train/Capture14’, ‘train/Capture15’, ‘train/Capture16’, ‘train/Capture20’, ‘train/Capture22’, ‘trai...

    2026

  45. [50]

    As shown in tab

    with [ 3, 41], as they show robust one-shot reconstruction on InterHand2.6M. As shown in tab. 3, our method outperform previous works consistently in all metrics. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Zhisheng Han et al. Figure 9: Points visibility and VPIA attention visualization. Figure 10: Robustness towards different input images. The i...

    2026

  46. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  47. [2020]

    6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image

    Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. InEuropean Conference on Computer Vision. Springer, 548–564

  48. [2023]

    Graph.42, 4 (2023), 139–1

    3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph.42, 4 (2023), 139–1

    2023

  49. [2024]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Authentic Hand Avatar from a Phone Scan via Universal Hand Model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2029–2038

    2029

  50. [2025]

    InProceedings of the Computer Vision and Pattern Recognition Conference

    Gst: Precise 3d human body from a single image with gaussian splatting transformers. InProceedings of the Computer Vision and Pattern Recognition Conference. 6007–6017