Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Gait can be recognized directly from a compact 184-dimensional vector of articulated body and hand controls, without silhouettes or dense meshes.

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 13:56 UTC pith:WQLMBPLJ

load-bearing objection Genuine, well-executed work on rig-space gait recognition, but the central claim needs a same-upstream control before it fully holds. the 3 major comments →

arxiv 2607.29083 v1 pith:WQLMBPLJ submitted 2026-07-31 cs.CV

MHRGait: Gait Recognition from Momentum Human Rig Pose

Huiran Duan , Qian Zhou , Xianda Guo , Hua Zou , Guoying Zhao , Zhongyuan Wang , Yingli Tian This is my paper
classification cs.CV
keywords gait recognitionarticulated poserig parametersbody and hand posemodel-based gaitclothing-change robustnessmultimodal fusiontemporal modeling
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.

This paper tries to show that the identity-bearing signal in walking survives even when all geometric clothing—silhouette shape, skeleton scale, surface mesh—is stripped away. Its proposed representation, MHR pose, describes each video frame by 76 body-pose and 108 hand-pose parameters recovered from monocular video, and the recognition network sees only these numbers. The paper argues that on cloth-changing benchmarks its standalone recognizer outperforms all compared skeleton- and silhouette-based model-based methods, and that merging MHR pose with silhouettes improves silhouette recognizers especially under clothing change and nighttime. Why it matters: if correct, gait recognition no longer needs to fight clothing and segmentation noise through shape cues; compact rig-space articulation is itself a competitive, transferable identity cue.

Core claim

On its own terms, the paper's discovery is that rig-space articulation is a sufficient and complementary gait representation. MHRGait groups the 184 per-frame controls into eight anatomical groups (six body, two hands), lets a Mixer-style token exchange model intra-frame coordination, and applies a lightweight temporal bottleneck to capture how the controls evolve over a walking cycle. The output is one body descriptor and one hand descriptor. Evaluated on four benchmarks, MHRGait reports the best overall model-based results on SUSTech1K (Rank-1 67.8%, Rank-5 86.9%) and CCPG (mean Rank-1 68.2%, up from 60.4%), and strong cross-dataset transfer; MHRGait++ adds silhouettes with modality-balanc

What carries the argument

The central object is the Momentum Human Rig (MHR) pose vector: a 184-dimensional per-frame control vector (76 body + 108 hand, 54 per hand) that describes articulation while global rotation, skeleton scale, shape, and expression are discarded. The argument is carried by three mechanisms: anatomy-aware semantic grouping (six body groups, two hand groups) with learnable projections per group; Mixer-style token mixing plus adaptive semantic aggregation to model intra-frame coordination; and a residual bottleneck with 1D temporal convolution to model gait dynamics. For the multimodal extension, the load-bearing mechanism is modality-balanced retrieval: distances are averaged within each modalit

Load-bearing premise

The load-bearing premise is that the frozen upstream tracker and monocular full-body estimator produce MHR pose controls that are accurate and anatomically consistent under clothing, occlusion, low light, and carrying; if upstream pose errors are large or identity-correlated, the recognition gains vanish.

What would settle it

Feed MHRGait ground-truth rig parameters from a motion-capture or synthetic dataset with fixed body shape and identity labels; if Rank-1 accuracy drops to chance, articulation alone does not carry identity and the reported results depend on estimator artifacts or shape leakage.

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

If this is right

  • Model-based gait recognition can operate from compact articulated controls alone, reducing the representation from dense meshes or silhouette maps to 184 numbers per frame.
  • A recognition network this small (2.76M parameters, 0.69 GFLOPs for 30 frames) can exceed heavier skeleton- and contour-based model-based baselines.
  • Rig-space articulation and projected body shape are complementary: adding MHR pose to a silhouette recognizer yields its largest gains under clothing change and nighttime, where shape cues fail.
  • Articulation cues transfer across datasets better than appearance/shape cues in the paper's experiments, suggesting the identity of gait dynamics is less dataset-specific.
  • Multimodal retrieval should balance modalities by information content, not by descriptor count; equal-part fusion hides a 2-descriptor modality behind a 16-descriptor one.

Where Pith is reading between the lines

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

  • A testable extension: train MHRGait on synthetic or motion-capture rig parameters with identical body shape and scale; if identity is still recoverable, the paper's core claim that articulation itself carries identity would be isolated from body-shape leakage.
  • The 12.7-point gain from adding hand controls to body-only modeling suggests fine-grained hand dynamics carry identity: this could transfer to close-range person identification or to medical gait analysis where subtle limb control matters.
  • The modality-balanced fusion principle generalizes: any multimodal retrieval system with uneven descriptor counts could adopt per-modality distance normalization before fusion.
  • Because the offline MHR construction pipeline costs on the order of 90 ms/frame and needs a tracker plus monocular estimator, practical deployment hinges on faster or end-to-end differentiable pose recovery; that is the paper's stated future direction, not its claim.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces MHR pose, a 184-D per-frame vector of body and hand articulation controls from the MHR parametric rig, as a gait recognition representation. It proposes MHRGait, a compact body-part-aware temporal encoder that groups the 184 controls into anatomical regions, models intra-frame coordination with MLP-Mixer blocks, and models temporal evolution with a residual temporal bottleneck. It further proposes MHRGait++, which fuses MHR pose with a DeepGaitV2 silhouette branch via modality-balanced distance fusion (Eq. 13). The method is evaluated on SUSTech1K, CCPG, CCGR-MINI, and CASIA-B*. The main empirical claims are: (i) MHRGait attains the best overall performance among compared model-based methods on CCPG and SUSTech1K (e.g., SUSTech1K Rank-1 67.8% vs. SkeletonGait's 50.1%); (ii) MHRGait++ consistently improves a strong silhouette recognizer, with large gains under clothing change and at night; and (iii) the recognition network is efficient (2.76M parameters, 0.69 GFLOPs for MHRGait). The authors conclude that rig-space articulation is an effective standalone gait representation and a complementary cue to projected body shape.

Significance. If the claims hold, the paper would make a useful contribution to model-based gait recognition: it demonstrates that a compact, semantically organized articulation space can rival or beat sparse joint coordinates and projected silhouettes, and that hand articulation carries identity information beyond body-only pose. The paper is also strong on engineering: it provides code, reports efficiency numbers (Table 12 and Supplementary Table 13), includes careful ablations of anatomical grouping and temporal modeling, and evaluates on four benchmarks with a realistic cross-domain protocol. The central scientific claim, however, is currently underdetermined because all skeleton baselines use different upstream pose estimators. The paper's contribution is novel relative to existing skeleton-based and SMPL-based gait methods, and the cross-domain gains are suggestive, but the representation-level superiority is not cleanly isolated without a same-upstream control.

major comments (3)
  1. [Experiments, Table 1; Fig. 2; Eqs. (1)-(2)] The central claim that MHR pose is an effective standalone gait representation is not cleanly isolated from upstream pose-estimation quality. MHRGait uses a frozen SAM3 tracker plus SAM 3D Body to obtain MHR controls, whereas every skeleton baseline in Table 1 (GaitGraph2, GaitTR, GPGait, GPGait++, SkeletonGait, GaitContour) was published with its own 2D/3D keypoint estimator. The reported gains over SkeletonGait (12.3 points Rank-1 on SUSTech1K, 7.8 points mean on CCPG) could therefore reflect the higher accuracy and temporal stability of the upstream estimator rather than the rig-space representation itself. The paper does not include any baseline that feeds the same SAM3/SAM 3D Body outputs, converted to Cartesian joints or mesh vertices, into a strong skeleton-based or mesh-based recognizer. Without this control, the conclusion that 'rig-space articulation... establishes' itself as a
  2. [Supplementary Table 10; Eq. (13)] The MHRGait++ fusion weight α=0.5 is selected by sweeping α on the CCPG and SUSTech1K test sets and choosing the value with the best accuracy. The text states that the sweep 'evaluates sensitivity to the retrieval weight rather than performing dataset-specific tuning,' but choosing the value that maximizes test accuracy is model selection on evaluation data, even if the same value is then used across datasets. This makes the headline MHRGait++ results (e.g., 90.0% mean Rank-1 on CCPG, 90.6/96.6% on SUSTech1K) optimistically biased. Please either select α on a held-out validation split, or present Table 10 strictly as a sensitivity analysis without adopting the best point as the reported configuration.
  3. [Tables 1-6] All reported numbers are single-run point estimates with no error bars, standard deviations, or multiple-seed statistics. This is particularly consequential for close comparisons: in Table 4, MHRGait++ and Origins-M differ by only 0.1 points in Rank-1 (41.6 vs. 41.5) and MHRGait++ has lower mAP (38.0 vs. 38.3), so the claim of 'highest Rank-1 accuracy on CCGR-MINI' is not supported without variance information. Please report results over multiple seeds (at least for the proposed methods and the most competitive baselines) or provide a statistical test for the key comparisons.
minor comments (5)
  1. [Table 1; Related Work] The related work cites parametric-model baselines SMPL-Gait (Zheng et al., CVPR'22) and VM-Gait (Wang et al., WACV'25), but neither appears in Table 1's model-based comparison. If these baselines are not directly comparable under the same protocol, say so explicitly; otherwise their omission weakens the 'best model-based' claim.
  2. [Eq. (13)] The notation D_S(V_p,V_g) and D_M(V_p,V_g) is not fully defined before Eq. (13). Please state explicitly that D_S and D_M are within-modality averages of part-wise Euclidean distances, and clarify whether the probe/gallery descriptors are L2-normalized before distance computation.
  3. [Supplementary Table 12; Table 13] The efficiency claims are for the recognition network only, excluding the offline MHR construction. Given that the full MHR pipeline is 91.6 ms/frame versus 12.4 ms/frame for silhouettes on CCPG, the paper's 'favorable accuracy-efficiency trade-off' is only about the recognition stage. Please add a sentence in the main text clarifying that the reported GFLOPs do not include pose recovery and tracking.
  4. [Supplementary Table 10] The row α=0 (silhouette-only after multimodal training) gives 82.3% Rank-1 on SUSTech1K, which is substantially lower than the standalone DeepGaitV2 number 87.4% in Table 2. Please explain why joint training with the MHR branch hurts the silhouette-only retrieval, or note that the two numbers are not directly comparable because of different training protocols.
  5. [General presentation] Figure captions and table headings contain typographical issues, e.g., 'gaitevaluationprotocol' in the Table 3 caption and inconsistent use of 'Rank-1' vs. 'R-1' across tables. Please proofread the final version.

Circularity Check

0 steps flagged

No significant circularity: MHRGait's central claims are benchmarked against external data and prior methods; no prediction reduces by construction to its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional: frozen SAM3 tracking and SAM 3D Body produce an MHR pose sequence (Eqs. 1-2), MHRGait encodes it into body/hand descriptors, and rank-1 accuracy is measured against external benchmarks (SUSTech1K, CCPG, CCGR-MINI, CASIA-B*) and compared with baselines from other groups. No quantity called a prediction is constructed from the same fitted quantity that defines it. The MHR representation is the input, not an output derived from identity labels. The MHRGait++ retrieval weight α=0.5 is a fixed hyperparameter, and the supplementary sweep over α evaluates sensitivity on the evaluation sets rather than fitting a constant that mathematically forces the reported accuracy; this is a model-selection concern, not constructional circularity. The only self-citation is a related-work reference to GaitFuse in the introduction, which is not load-bearing for the central claim. The conclusion's admitted limitation that upstream pose-estimation errors may propagate is a correctness/robustness risk, not circularity. No uniqueness theorem from the authors is invoked, no ansatz is smuggled in via citation, and no known result is merely renamed as a new derivation. The central empirical claim is independent of any internal fitted parameter that defines the target metric, so the paper is self-contained against external benchmarks and receives score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on an offline pose-estimation pipeline whose reliability is assumed, plus several hand-chosen grouping and hyperparameter choices. No new physical entities are introduced. The empirical benchmarks are external, so the ledger is dominated by domain assumptions about upstream MHR estimation and baseline comparability.

free parameters (4)
  • MHRGait++ retrieval weight α = 0.5
    Swept in Supp. Table 10; 0.5 gives the best CCPG/SUSTech1K accuracy and is fixed across datasets; no separate validation split is reported.
  • Anatomical group boundaries = Indices per Supp. Table 14 (8 groups)
    Hand-assigned partition of the 184-D MHR vector into torso/head, arms, legs/feet, flexibility, hands; central to the semantic encoding and not learned.
  • Encoder hyperparameters = d_h=d=256, L=4 mixer blocks, temporal kernel=9, reduction ratio=4
    Manual architecture choices; ablations show grouping matters, but no sensitivity analysis is given for these dimensions.
  • Training schedules = lr=0.1, momentum=0.9, weight decay=5e-4, per-dataset milestones/iterations
    Per-dataset settings in Supp. Table 9; standard but tuned, with no sensitivity study.
axioms (4)
  • domain assumption Frozen SAM3 tracking plus SAM 3D Body yield MHR pose controls accurate enough for identity recognition.
    Fig. 2 and Eqs. (1)-(2) treat the estimated MHR pose sequence as reliable input; the Conclusion concedes upstream errors may propagate.
  • domain assumption MHR pose controls are decoupled from clothing/body shape, so articulation survives clothing changes.
    Central motivation in the Introduction and Figure 5, but the paper does not directly measure decoupling quality under covariates.
  • domain assumption Reported external baseline numbers and official dataset splits/protocols are accurate and comparable.
    Tables 1-6 rely on published baselines; CASIA-B* is a re-segmented version, and baseline numbers are not reimplemented.
  • domain assumption Body+hand articulation trajectories are identity-discriminative in these benchmarks.
    Ablation Table 8 shows body alone is 59.8 R-1 on SUSTech1K and hand alone 25.4, so the joint signal carries the claim.

pith-pipeline@v1.3.0-daily-deepseek · 20661 in / 13214 out tokens · 144627 ms · 2026-08-03T13:56:34.258310+00:00 · methodology

0 comments
read the original abstract

Gait recognition is shaped by its input representation. Silhouettes encode projected body shape, skeletons encode sparse joint coordinates, and 3D meshes encode dense surface geometry. In each case, identity-bearing articulation is observed through geometric carriers that also vary with clothing, skeletal scale, or body shape. We investigate whether gait can instead be recognized from compact articulated controls. We introduce Momentum Human Rig (MHR) pose as a gait representation, describing each frame using 184 semantically organized body and hand parameters estimated from monocular video. MHRGait groups these heterogeneous controls by anatomy, models their intra-frame coordination and temporal evolution, and produces compact body and hand descriptors. We further introduce MHRGait++, which combines MHR pose with silhouettes through modality-balanced distance fusion, preventing descriptor count from determining modality importance. Experiments on four benchmarks show that MHRGait attains the best overall performance among compared model-based methods on CCPG and SUSTech1K and transfers effectively across datasets, while its recognition network requires only 2.76M parameters and 0.69 GFLOPs for a 30-frame input. MHRGait++ consistently improves silhouette recognizers with a favorable accuracy-efficiency trade-off. These results establish rig-space articulation as an effective standalone gait representation and a complementary cue to projected body shape. Our code is available at https://github.com/duanhuiran/MHRGait.

Figures

Figures reproduced from arXiv: 2607.29083 by Guoying Zhao, Hua Zou, Huiran Duan, Qian Zhou, Xianda Guo, Yingli Tian, Zhongyuan Wang.

Figure 1
Figure 1. Figure 1: Overview of gait representations and their [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Offline construction of an MHR pose sequence. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of MHRGait, which encodes body and hand articulation into a compact two-part gait descriptor. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Modality-balanced retrieval in MHRGait++. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representation comparison on CCPG under clothing change. Silhouettes, Cartesian body representations, decoded MHR pose, and the left-knee bend control are shown for two clothing trials of the same identity. RGB video 40 45 50 55 60 15 20 25 30 35 sub-0639_seq-01-cr-nt_view-270 Silhouette SkeletonMap Skeleton MHR pose (decoded) Left hand part of decoded MHR sub-0604_seq-01-cr-cl_view-045 (a) Clothing + Carr… view at source ↗
Figure 6
Figure 6. Figure 6: Representation comparison on SUSTech1K under compound covariates. Silhouette, Cartesian-body, and decoded MHR representations are shown for clothing-plus-carrying and nighttime-plus-carrying sequences; the last row visualizes decoded left-hand pose [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Subject-level MHR articulation dynamics on SUSTech1K. Gait-cycle periods and selected MHR-control trajectories for identities 0000–0009. while the peak GPU memory decreases from 18.55/18.55 to 12.87/12.87 GB. Although full MHR construction is more expensive than silhouette or skeleton preprocessing, these results show that approximately half of its runtime is associ￾ated with the one-time tracking and mask… 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

49 extracted references · 1 linked inside Pith

  1. [1]

    IEEE Transactions on Multimedia , year=

    Watch where you move: Region-aware dynamic aggregation and excitation for gait recognition , author=. IEEE Transactions on Multimedia , year=

  2. [2]

    Gait Recognition in the Wild: A Large-Scale Benchmark and NAS-Based Baseline , year=

    Guo, Xianda and Zhu, Zheng and Yang, Tian and Lin, Beibei and Huang, Junjie and Deng, Jiankang and Huang, Guan and Zhou, Jie and Lu, Jiwen , journal=. Gait Recognition in the Wild: A Large-Scale Benchmark and NAS-Based Baseline , year=

  3. [3]

    DyGait: Exploiting Dynamic Representations for High-performance Gait Recognition , year=

    Wang, Ming and Guo, Xianda and Lin, Beibei and Yang, Tian and Zhu, Zheng and Li, Lincheng and Zhang, Shunli and Yu, Xin , booktitle=. DyGait: Exploiting Dynamic Representations for High-performance Gait Recognition , year=

  4. [4]

    2021 IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Gait recognition in the wild: A benchmark , author=. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    2021

  5. [5]

    Information Fusion , volume =

    GaitFuse: A Hierarchical Cross-Modal Alignment and Uncertainty-Aware Fusion Framework for Multi-Modal Gait Recognition , author=. Information Fusion , volume =. 2026 , publisher=

    2026

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Lidargait: Benchmarking 3d gait recognition with point clouds , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [8]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Cross-Covariate Gait Recognition: A Benchmark , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi =

    2024

  9. [9]

    A Framework for Evaluating the Effect of View Angle, Clothing and Carrying Condition on Gait Recognition , year=

    Shiqi Yu and Daoliang Tan and Tieniu Tan , booktitle=. A Framework for Evaluating the Effect of View Angle, Clothing and Carrying Condition on Gait Recognition , year=

  10. [10]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Gaitset: Regarding gait as a set for cross-view gait recognition , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  11. [11]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Gaitpart: Temporal part-based model for gait recognition , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  12. [12]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Gait recognition via effective global-local feature representation and local temporal aggregation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  13. [13]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Opengait: Revisiting gait recognition towards better practicality , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [14]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Opengait: A comprehensive benchmark study for gait recognition towards better practicality , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  15. [15]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Learning a unified template for gait recognition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  16. [16]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Gait Transformer: End-to-End Transformer Backbone for Gait Recognition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  17. [17]

    Towards a Deeper Understanding of Skeleton-based Gait Recognition , year=

    Teepe, Torben and Gilg, Johannes and Herzog, Fabian and Hörmann, Stefan and Rigoll, Gerhard , booktitle=. Towards a Deeper Understanding of Skeleton-based Gait Recognition , year=

  18. [18]

    Expert Systems , volume=

    Spatial transformer network on skeleton-based gait recognition , author=. Expert Systems , volume=. 2023 , publisher=

    2023

  19. [19]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Gpgait: Generalized pose-based gait recognition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  20. [20]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Skeletongait: Gait recognition using skeleton maps , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  21. [21]

    From FastPoseGait to GPGait++: Bridging the Past and Future for Pose-Based Gait Recognition , year=

    Meng, Shibei and Fu, Yang and Hou, Saihui and Hu, Xuecai and Cao, Chunshui and Liu, Xu and Huang, Yongzhen , journal=. From FastPoseGait to GPGait++: Bridging the Past and Future for Pose-Based Gait Recognition , year=

  22. [22]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages =

    Zheng, Jinkai and Liu, Xinchen and Zhang, Boyue and Yan, Chenggang and Zhang, Jiyong and Liu, Wu and Zhang, Yongdong , title =. Proceedings of the 32nd ACM International Conference on Multimedia , pages =. 2024 , isbn =

    2024

  23. [23]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    GaitContour: Efficient Gait Recognition Based on a Contour-Pose Representation , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  24. [24]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Exploring more from multiple gait modalities for human identification , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  25. [25]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    On Denoising Walking Videos for Gait Recognition , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  26. [26]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Biggait: Learning gait representation you want by large vision models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  27. [27]

    Advances in neural information processing systems , volume=

    Biggergait: Unlocking gait recognition with layer-wise representations from large vision models , author=. Advances in neural information processing systems , volume=

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unlocking motion from large vision models with a semantic and kinematic duality for gait recognition , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Gait-X: Exploring x modality for generalized gait recognition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [30]

    European Conference on Computer Vision , pages=

    Gaitedge: Beyond plain end-to-end gait recognition for better practicality , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  31. [31]

    MLP-Mixer: An all-MLP Architecture for Vision , volume =

    Tolstikhin, Ilya O and Houlsby, Neil and Kolesnikov, Alexander and Beyer, Lucas and Zhai, Xiaohua and Unterthiner, Thomas and Yung, Jessica and Steiner, Andreas and Keysers, Daniel and Uszkoreit, Jakob and Lucic, Mario and Dosovitskiy, Alexey , booktitle =. MLP-Mixer: An all-MLP Architecture for Vision , volume =

  32. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , month =

    Luo, Hao and Gu, Youzhi and Liao, Xingyu and Lai, Shenqi and Jiang, Wei , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , month =

  33. [33]

    , journal =

    Loper, Matthew and Mahmood, Naureen and Romero, Javier and Pons-Moll, Gerard and Black, Michael J. , journal =. 2015 , doi =

    2015

  34. [34]

    ACM Transactions on Graphics , volume =

    Embodied Hands: Modeling and Capturing Hands and Bodies Together , author =. ACM Transactions on Graphics , volume =

  35. [35]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Expressive Body Capture: 3D Hands, Face, and Body from a Single Image , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  36. [36]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Monocular, One-Stage, Regression of Multiple 3D People , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  37. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  38. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Humans in 4D: Reconstructing and Tracking Humans with Transformers , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  39. [39]

    Park, Jinhyung and Romero, Javier and Saito, Shunsuke and Prada, Fabian and Shiratori, Takaaki and Xu, Yichen and Bogo, Federica and Yu, Shoou-I and Kitani, Kris and Khirodkar, Rawal , booktitle =

  40. [40]

    Ferguson, Aaron and Osman, Ahmed A. A. and Bescos, Berta and Stoll, Carsten and Twigg, Chris and Lassner, Christoph and Otte, David and Vignola, Eric and Prada, Fabian and Bogo, Federica and others , journal =

  41. [41]

    2026 , pages =

    Yang, Xitong and Kukreja, Devansh and Pinkus, Don and Fan, Taosha and Park, Jinhyung and Shin, Soyong and Cao, Jinkun and Liu, Jia-Wei and Ugrinovic, Nicol\'as and Sagar, Anushka and Malik, Jitendra and Feiszli, Matt and Doll\'ar, Piotr and Kitani, Kris , booktitle =. 2026 , pages =

    2026

  42. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Gait Recognition in the Wild with Dense 3D Representations and a Benchmark , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  43. [43]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =

    End-to-End Model-Based Gait Recognition Using Synchronized Multi-View Pose Constraint , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =. 2021 , pages =

    2021

  44. [44]

    Wang, Zhao Yang and Liu, Jiang and Chen, Jieneng and Chellappa, Rama , booktitle =

  45. [45]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Multi-modal gait recognition via effective spatial-temporal feature fusion , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  46. [46]

    2511.16719 , archivePrefix =

    Carion, Nicolas and Gustafson, Laura and Hu, Yuan-Ting and Debnath, Shoubhik and Hu, Ronghang and Suris, Didac and Ryali, Chaitanya and Alwala, Kalyan Vasudev and Khedr, Haitham and Huang, Andrew and Lei, Jie and Ma, Tengyu and Guo, Baishan and Kalla, Arpit and Marks, Markus and Greer, Joseph and Wang, Meng and Sun, Peize and Rädle, Roman and Afouras, Tri...

    Pith/arXiv arXiv

  47. [47]

    2025 , url =

    Gao, Mingqi and Miao, Yunqi and Han, Jungong , journal =. 2025 , url =

    2025

  48. [48]

    2024 , note =

    Ultralytics YOLO11 , author =. 2024 , note =

    2024

  49. [49]

    2022 , pages =

    Chu, Lutao and Liu, Yi and Wu, Zewu and Tang, Shiyu and Chen, Guowei and Hao, Yuying and Peng, Juncai and Yu, Zhiliang and Chen, Zeyu and Lai, Baohua and Xiong, Haoyi , booktitle =. 2022 , pages =

    2022