Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that intermediate layers of large vision models carry more discriminative gait information than the final layer, and that averaging per-layer distance scores yields state-of-the-art gait recognition without heavy gait…

desk verdict A simple and useful baseline for LVM-based gait recognition; the per-layer loss specification is unclear, but the central fusion claim is supported and worth peer review. read the letter →

arxiv 2505.18132 v3 pith:LV7F3UKH submitted 2025-05-23 cs.CV

Dingqiang Ye , Chao Fan , Zhanbo Huang , Chengwen Luo , Jianqiang Li , Shiqi Yu , Xiaoming Liu This is my paper · ORCID
classification cs.CV
keywords gaitrecognitionlargevisionmodelslayer-wiserepresentationsDINOv2CLIPSAMensembletestingcross-domainevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

BiggerGait tests the hypothesis that the value of large vision models for gait recognition lies in their internal layer diversity rather than in hand-crafted gait priors. The paper shows that across DINOv2, CLIP, and SAM, the best Rank-1 accuracy tends to come from intermediate layers, not the final layer, and that averaging per-layer distance scores at inference yields substantial gains. It builds a simple layer-wise baseline on this observation and reports state-of-the-art results on the CCPG benchmark and in cross-domain evaluation. A grouped variant shares one gait encoder across layers, preserving most of the gain while cutting parameters and FLOPs.

What carries the argument

The central object is the set $R$ of per-layer gait embeddings $g_i$ produced by per-layer linear projections and GaitBase heads, together with the ensemble similarity $D(x,y)$ defined as the mean of the per-layer Euclidean distances $d_i(x,y)$. The core mechanism is layer-wise complementarity: none of the single layers is consistently best, but averaging all per-layer scores usually beats every individual layer. The paper's grouped variant, BiggerGait*, clusters adjacent layers into $J$ depth groups and shares $P$ gait encoders, trading a small accuracy loss for fewer parameters and FLOPs.

What would settle it

Retrain BiggerGait with the same frozen LVM features and silhouette mask but a single shared head, then feed each layer through that head at test time; if the middle-layer accuracy peak disappears or sharply shrinks, the middle-layer advantage is an artifact of per-layer training rather than a property of the frozen features.

Watch

Extended reading notes

Core claim

The discovery, stated on the paper's own terms, is that large vision models carry complementary gait information spread across their layers. On CCPG, the best middle layer of DINOv2-S reaches 89.5% Rank-1 while the final layer drops to 80.9%, and mean-based ensemble testing across twelve layers adds roughly +7.7%, +10.3%, and +7.0% Rank-1 for DINOv2-S, CLIP-S, and SAM-S. From this the paper derives BiggerGait, which samples twelve layers, removes background with an auto-encoder silhouette mask, projects each masked feature, runs a separate GaitBase head per layer, and fuses by averaging distances. With this recipe it reports state-of-the-art within-domain results on CCPG and strong cross-domain transfer to CASIA-B*, SUSTech1K, and CCGR_MINI.

Load-bearing premise

The layer-wise conclusions depend on the unstated assumption that the training losses are applied to each layer's representation separately, so that the reported per-layer accuracies are clean measurements of each LVM layer; if the losses act only on a fused representation, those curves are not what they appear to be.

Editorial extensions

If this is right

  • BiggerGait's results imply that heavy human-designed gait priors are not necessary for strong LVM-based gait recognition; a simple auto-encoder silhouette mask suffices.
  • Mean-based ensemble testing across twelve layers is an effective test-time fusion, raising Rank-1 by up to 10.3%.
  • The best layer for gait varies by dataset, so layer-wise evaluation rather than final-layer evaluation should be standard for RGB gait recognition.
  • For large LVM variants the standard BiggerGait is recommended, while the grouped variant BiggerGait* provides most of the gain at lower cost for smaller models.
  • The data-bias limitation remains: training on CCGR_MINI transfers to CCPG unevenly, and within-domain gains on CCGR_MINI are modest.

Reading between the lines

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

  • A reader should be cautious that the per-layer accuracy curves measure each trained per-layer head plus its frozen LVM layer, not the frozen layer alone; rerunning with a single shared head would show how much of the middle-layer advantage is intrinsic.
  • The same ensemble-of-per-layer-distances recipe may transfer to other fine-grained recognition tasks where appearance covariates vary, such as person re-identification, but this is untested.
  • The grouping result suggests a cheap diagnostic: layer representations share enough structure that one or two trained heads suffice; that prediction can be checked on other ViT backbones without retraining the full pipeline.
  • The CCGR_MINI case where fusion hurts suggests that when layers are not complementary a fixed mean is suboptimal; a learnable per-layer weighting could be explored.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper claims that large vision models (LVMs) contain rich, layer-wise gait-relevant representations and that intermediate layers are more discriminative than the final layer, task-dependent, and complementary. It proposes BiggerGait, a baseline that samples 12 LVM layers, applies a learned silhouette mask (the only explicit gait prior), per-layer linear projections and GaitBase extractors, and fuses per-layer distances by averaging at inference. A grouped variant BiggerGait* shares gait encoders across depth groups to reduce cost. Experiments on CCPG, CCGR_MINI, CASIA-B*, and SUSTech1K report state-of-the-art or competitive results across DINOv2, CLIP, and SAM backbones in within- and cross-domain settings.

Significance. If the results hold, the paper offers a simple and surprisingly effective recipe: fuse intermediate LVM layers with lightweight per-layer heads and minimal gait priors, rather than designing task-specific losses or heavy domain knowledge. The main evidence is the consistent middle-layer advantage across three architecturally different LVMs and four datasets, plus the ensemble gains in Fig. 5. The authors also provide an efficiency-oriented grouped variant and report detailed ablations. The paper ships no code in the manuscript, but the proposed method is simple enough that the missing training details are the key obstacle to reproduction and to interpreting the central claims.

major comments (4)
  1. [Sec. 3.1-3.2, Eq. (5)] The manuscript never specifies how the triplet and cross-entropy losses in Eq. (5) are attached to the per-layer embeddings g_i in Eq. (4). The separate-testing protocol in Sec. 3.2 and the layer-wise accuracy curves in Figs. 3-4 are interpretable as per-layer measurements only if L_tri and L_ce are applied to each g_i individually (e.g., with per-layer classification heads and triplet sampling). If the losses act only on a fused representation, those curves are not clean per-layer measurements; if they act per layer, the middle-layer advantage may partly reflect the training objective. Please state the exact loss composition, including the number of classifiers and the triplet mining strategy, and explain how this supports the claims in Secs. 3.3-3.5.
  2. [Tables 1-2 and Figs. 3-4] No measure of variance is reported anywhere. Tables 1-2 and Figs. 3-4 give single Rank-1 numbers, yet several conclusions rest on small differences (e.g., the 83.3% probability that the best cross-domain layer differs from the best within-domain layer, and the choice of J=6 and P=2 over other configurations in Table 2(b)). Without standard deviations or multiple seeds, those differences cannot be distinguished from optimization noise. Please report means and standard deviations over at least three runs for the layer-wise curves, the main comparison tables, and the ablation table.
  3. [Sec. 3.2 and Sec. 4.1] The layer sampling and backbone training status are underspecified. Sec. 3.2 says the authors 'test every one of the 12 layers in LVM-S' but 'uniformly sample 12 layers from LVM-L'; the exact layer indices for LVM-L are never given. In addition, it is never stated whether the LVM weights are frozen or fine-tuned during training. This matters both for reproduction and for the interpretation of the 'intrinsic value of LVM itself' claim, since fine-tuning would let the training objective reshape the layer-wise representations. Please specify the sampled layer indices and the frozen/fine-tuned status of each backbone.
  4. [Sec. 3.5, Eq. (7)] Eq. (7) averages raw per-layer Euclidean distances d_i(x,y) without any normalization or scale calibration. If the per-layer embeddings have very different magnitudes, the ensemble score will be dominated by the layers with the largest distances, and the reported ensemble gains in Fig. 5(c) could be an artifact of scale rather than of complementarity. Please state whether the embeddings are normalized before computing d_i and whether any weighting or calibration is applied in the ensemble.
minor comments (5)
  1. [Sec. 3.1, Eq. (1)] The mask m is obtained by applying softmax along the channel dimension to a 2-channel output E(f_N), but Eq. (2) multiplies this mask with f_i. Please clarify whether m is a single-channel probability map or a 2-channel multiplicative mask, and how the channel dimensions align.
  2. [Sec. 3.1] There is a typo: 'lightwight' should be 'lightweight'.
  3. [Tables 2(a) and 2(b)] The heading 'Ablation of Gait Group (J=12)' and 'Ablation of Depth Group (P=2,2,3)' is confusing; please explicitly define the value of P and J in each row and clarify that Table 2(b) uses different P values per LVM.
  4. [Table 1] The CASIA-B* columns are labeled 'CL', 'NM', and 'BG' but the standard protocol has NM, BG, and CL as clothing/carrying conditions; please state the evaluation protocol and the number of gallery/probe views used.
  5. [Sec. 5] The limitation statement says temporal features are underexplored, but Sec. 4.1 describes 30-frame sequences with GaitBase temporal aggregation; please clarify that the temporal modeling is inherited from GaitBase rather than newly studied.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the layer-wise measurements and ensemble gains are empirical outputs obtained against external benchmarks, not quantities defined into existence by the paper's own equations.

full rationale

The paper's central claims—middle layers outperform final layers and ensembling per-layer distances improves accuracy—are supported by controlled experiments on external benchmarks (CCPG, CASIA-B*, SUSTech1K, CCGR_MINI) rather than by construction. Equation (7) is a plain average of per-layer distances and is compared with separate testing; no fitted parameter is renamed as a prediction. The only self-citations are to the authors' prior BigGait [54] for the silhouette auto-encoder design and the channel width C=16; these are implementation borrowings, not load-bearing justifications of the layer-complementarity result, and BigGait is a published CVPR artifact external to this paper's fitted values. The grouping strategy cites external work [46] for the shared-feature-space hypothesis, not a self-citation. The one flagged concern is that Eq. (5) does not explicitly state whether triplet and cross-entropy losses are applied per layer or to a fused representation; this is a reproducibility/attribution ambiguity, not a circular step, because the reported per-layer accuracies are empirical outputs of a trained system and the paper does not define the middle-layer advantage into existence through the loss. The Limitation section's note about temporal features being underexplored is a scope limitation, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. Its scientific claim rests on domain assumptions about transferability of frozen LVM features, the validity of a shared feature space for adjacent layers, and the correctness of a single silhouette mask applied to all layers. Hand-chosen hyperparameters include N=12, C=16, J=6, and P=2. The unspecified input to the triplet and cross-entropy losses is a gap rather than a fitted parameter.

free parameters (4)
  • N = 12 sampled LVM layers = 12
    Number of layers sampled from each LVM, chosen by design and not derived. For LVM-S this is all 12 layers; for LVM-L the 12 layers are uniformly sampled without listing the exact indices.
  • Projection output channels C = 16
    Output channel of the lightweight linear projection, set to 16 following BigGait [54].
  • Number of depth groups J = 6
    Selected via ablation on the CCPG benchmark in Table 2(b) and used for the BiggerGait* results in Table 1. Tuning is done on the same benchmark used for reporting.
  • Number of gait groups P = 2
    Selected via ablation on CCPG in Table 2(a); reduces the number of shared gait encoders from 12 to 2.
assumptions (5)
  • domain assumption Pretrained CLIP, SAM, and DINOv2 features transfer to gait identity recognition despite never being trained on gait.
    The entire method relies on frozen LVM features containing identity-relevant cues. The paper tests this empirically but does not prove it.
  • domain assumption Middle-layer findings from LLMs, where intermediate layers are more robust or transferable, carry over to large vision models.
    Used to motivate the layer-wise analysis in Section 3. The experiments support it for gait, but it is not a mathematical guarantee.
  • domain assumption Residual connections make adjacent LVM layers share a feature space, so one shared gait encoder can process grouped layers without much accuracy loss.
    Used to justify the BiggerGait* grouping strategy in Section 3.6, based on reference [46] and on Figure 6(a). If the shared space is only approximate, the grouped variant could lose accuracy.
  • domain assumption The silhouette mask m = softmax(E(f_N)) correctly separates the person from the background in every layer, including shallow layers.
    All layer features are masked with the same mask derived from the final layer in Eq. (1). An inaccurate mask could distort lower-level features and bias the layer ranking.
  • standard math Triplet loss and cross-entropy loss are valid training signals for gait embeddings.
    Standard metric-learning losses adopted from prior gait work and not derived in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models." pith.science (2026). https://pith.science/paper/LV7F3UKH

@misc{pith2026250518132,
  author       = {Pith},
  title        = {Pith review of: BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LV7F3UKH}},
  note         = {Machine review of arXiv:2505.18132}
}
read the original abstract

Large vision models (LVM) based gait recognition has achieved impressive performance. However, existing LVM-based approaches may overemphasize gait priors while neglecting the intrinsic value of LVM itself, particularly the rich, distinct representations across its multi-layers. To adequately unlock LVM's potential, this work investigates the impact of layer-wise representations on downstream recognition tasks. Our analysis reveals that LVM's intermediate layers offer complementary properties across tasks, integrating them yields an impressive improvement even without rich well-designed gait priors. Building on this insight, we propose a simple and universal baseline for LVM-based gait recognition, termed BiggerGait. Comprehensive evaluations on CCPG, CAISA-B*, SUSTech1K, and CCGR\_MINI validate the superiority of BiggerGait across both within- and cross-domain tasks, establishing it as a simple yet practical baseline for gait representation learning. All the models and code will be publicly available.

Figures

Figures reproduced from arXiv: 2505.18132 by the authors.

Figure 1
Figure 1. Comparison of gait representation paradigms. (a) Pre-defined gait learning uses explicit gait patterns for prediction, losing essential identity information. (b) Typical LVM gait learning relies heavily on gait priors and treats all LVM layers equally, underutilizing discriminative features. (c) Our layer-wise LVM gait learning fully utilizes intermediate LVM layers with minimal gait priors. is this dependence truly… view at source ↗
Figure 2
Figure 2. Overview of the proposed BiggerGait. An LVM extracts multi-level features from RGB videos. Human silhouettes are generated using an unsupervised auto-encoder [54], serving as the only human-designed gait prior. Each level’s features, with background noise removed, are processed by separate linear projection layers and gait extractors to obtain the final gait representations. where E and D are 1×1 convolution layers,… view at source ↗
Figure 3
Figure 3. Layer-wise Performance Across LVMs. This figure presents the gait recognition accuracy of six large vision models, i.e., DINOv2-Small/Large [34], CLIP-Small/Large [38], and SAM-Small/Large [21], across 12 intermediate layers, evaluated on the CCPG [23] dataset. Each cell lists the Rank-1 accuracy for full clothing (CL), top (UP), pants (DN), and bag (BG) changes, along with the average (AVG). A downward arrow on the… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Layer-wise Performance Across Datasets. The columns (A, B, C, D) represent the four test sets, CCPG [23], SUSTech1K [40], CASIA-B* [55], and CCGR_MINI [59]. All models are trained on CCPG dataset. Column A reports within-domain performance, whereas columns B–D present …
Figure 5
Figure 5. Figure 5: Separate vs. Ensemble Layer Testing (a) The red box highlights the peak performance achieved by these layers. This peak performance represents the final result of separate testing. (b) Ensemble testing works likes a fair voting manner. (c) All models are trained on CCP…
Figure 6
Figure 6. Figure 6: (a) Average pairwise cosine similarity across the 12 layers of DINOv2-S [34], with depth and gait groups highlighted. (b) BiggerGait*: A grouped based BiggerGait that clusters similar layers (J = 6), uses two shared gait encoders (P = 2) to replace original per-layer o…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Decoding Children's Gait Behavior

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A new 1,185-video pediatric gait dataset with EVGS labels, plus a VideoMAE-based model, reaches ~84% average accuracy on 34 clinical gait items — far above MLLMs and prior gait models.

  2. GaitFace: A Multimodal Dataset for Long-Range Person Identification

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A 70-subject public long-range face+gait dataset and protocols show SOTA models collapse on native low-res and elevated 100 m probes despite strong optical-zoom performance.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages · cited by 2 Pith papers

  1. [1]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language Models are Few-Shot Learners. InAdvances in Neural Information Processing Systems (NeurIPS), pages 1877–1901, 2020. 3

    work page 1901

  2. [2]

    Cao and A

    K. Cao and A. K. Jain. Automated latent fingerprint recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41(4):788–800, 2018. 1

    work page 2018

  3. [3]

    Z. Cao, T. Simon, S.-E. Wei, and Y . Sheikh. Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. InComputer Vision and Pattern Recognition (CVPR), pages 7291–7299, 2017. 1, 3

    work page 2017

  4. [4]

    H. Chao, K. Wang, Y . He, J. Zhang, and J. Feng. Gaitset: Cross-view gait recognition through utilizing gait as a deep set.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(7):3467–3478,

    work page

  5. [5]

    L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. InEuropean Conference on Computer Vision (ECCV), pages 801–818, 2018. 9

    work page 2018

  6. [6]

    N. Chen, N. Wu, S. Liang, M. Gong, L. Shou, D. Zhang, and J. Li. Is bigger and deeper always better? probing llama across scales and layers.arXiv preprint arXiv:2312.04333, 2023. 2, 3, 5

    arXiv 2023

  7. [7]

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690–4699, 2019. 1

    work page 2019

  8. [8]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018. 3

    arXiv 2018

Show all 60 references
  1. [9]

    C. Fan, Y . Peng, C. Cao, X. Liu, S. Hou, J. Chi, Y . Huang, Q. Li, and Z. He. Gaitpart: Temporal part-based model for gait recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14225–14233, 2020. 1, 3, 8

    2020

  2. [10]

    C. Fan, S. Hou, Y . Huang, and S. Yu. Exploring Deep Models for Practical Gait Recognition.arXiv preprint arXiv:2303.03301, 2023. 8

    2023 arXiv

  3. [11]

    C. Fan, J. Liang, C. Shen, S. Hou, Y . Huang, and S. Yu. Opengait: Revisiting gait recognition towards better practicality. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9707–9716, 2023. 8

    2023

  4. [12]

    C. Fan, S. Hou, J. Liang, C. Shen, J. Ma, D. Jin, Y . Huang, and S. Yu. Opengait: A comprehensive benchmark study for gait recognition towards better practicality.arXiv preprint arXiv:2405.09138, 2024. 1, 3, 4

    2024 arXiv

  5. [13]

    C. Fan, J. Ma, D. Jin, C. Shen, and S. Yu. Skeletongait: Gait recognition using skeleton maps. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 1662–1669, 2024. 1, 3, 8

    2024

  6. [14]

    S. Fan, X. Jiang, X. Li, X. Meng, P. Han, S. Shang, A. Sun, Y . Wang, and Z. Wang. Not all layers of llms are necessary during inference.arXiv preprint arXiv:2403.02181, 2024. 2, 3, 5

    2024 arXiv

  7. [15]

    Y . Guo, S. Huang, R. Prabhakar, C. P. Lau, R. Chellappa, and C. Peng. Distillation-guided representation learning for unconstrained gait recognition. In2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–11. IEEE, 2024. 3

    2024

  8. [16]

    Y . Guo, A. Shah, J. Liu, A. Gupta, R. Chellappa, and C. Peng. Gaitcontour: Efficient gait recognition based on a contour-pose representation. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1051–1061. IEEE, 2025. 3

    2025

  9. [17]

    A. K. Jain, A. A. Ross, K. Nandakumar, and T. Swearingen. Additional biometric traits. InIntroduction to Biometrics, pages 245–287. Springer, 2024. 1

    2024

  10. [18]

    D. Jin, C. Fan, W. Chen, and S. Yu. Exploring more from multiple gait modalities for human identification. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 2, 8

    2025

  11. [19]

    D. Jin, C. Fan, J. Ma, J. Zhou, W. Chen, and S. Yu. On denoising walking videos for gait recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page in press, 2025. 1, 3, 4 11

    2025

  12. [20]

    M. Kim, Y . Su, F. Liu, A. Jain, and X. Liu. Keypoint relative position encoding for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 244–255,

  13. [21]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Dollár, and R. Girshick. Segment anything, 2023. 2, 3, 4, 5, 8, 9

    2023

  14. [22]

    P. Li, Y . Xu, Y . Wei, and Y . Yang. Self-correction for human parsing.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260–3271, 2020. 9

    2020

  15. [23]

    W. Li, S. Hou, C. Zhang, C. Cao, X. Liu, Y . Huang, and Y . Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13824–1383...

    2023

  16. [24]

    X. Li, Y . Makihara, C. Xu, Y . Yagi, S. Yu, and M. Ren. End-to-end Model-based Gait Recognition. In Asian Conference on Computer Vision (ACCV), page [no pagination], 2020. 1, 3

    2020

  17. [25]

    Liang, C

    J. Liang, C. Fan, S. Hou, C. Shen, Y . Huang, and S. Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. InComputer Vision – ECCV 2022, 2022. 8

    2022

  18. [26]

    R. Liao, C. Cao, E. B. Garcia, S. Yu, and Y . Huang. Pose-based temporal-spatial network (ptsn) for gait recognition with carrying and clothing variations. InChinese conference on biometric recognition, pages 474–483. Springer, 2017. 1, 3

    2017

  19. [27]

    B. Lin, S. Zhang, and X. Yu. Gait recognition via effective global-local feature representation and local temporal aggregation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14648–14656, 2021. 1, 3, 8

    2021

  20. [28]

    F. Liu, M. Kim, A. Jain, and X. Liu. Controllable and guided face synthesis for unconstrained face recognition. InProceedings of the European Conference on Computer Vision (ECCV), pages 701–719. Springer, 2022. 1

    2022

  21. [29]

    F. Liu, R. Ashbaugh, N. Chimitt, N. Hassan, A. Hassani, A. Jaiswal, M. Kim, Z. Mao, C. Perry, Z. Ren, et al. Farsight: A physics-driven whole-body biometric system at large distance and altitude. InIEEE Winter Conference on Applications of Computer Vision, pages 6227–6236, 2024. 1

    2024

  22. [30]

    K. Liu, O. Choi, J. Wang, and W. Hwang. CDGNet: Class Distribution Guided Network for Human Parsing. InComputer Vision and Pattern Recognition (CVPR), pages 4473–4482, 2022. 1, 3

    2022

  23. [31]

    Loper, N

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A Skinned Multi-Person Linear Model.ACM Transactions on Graphics, 34(6):[no pagination], 2015. 1, 3

    2015

  24. [32]

    J. Ma, D. Ye, C. Fan, and S. Yu. Pedestrian attribute editing for gait recognition and anonymization.arXiv preprint arXiv:2303.05076, 2023. 4

    2023 arXiv

  25. [33]

    M. S. Nixon and J. N. Carter. Automatic Recognition by Gait.Proceedings of the IEEE, 94(11):2013–2024,

    2013

  26. [34]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision.arXiv preprint arXiv:2304.07193, 2023. 2, 3, 4, 5, 6, 7, 8, 9

    2023 arXiv

  27. [35]

    Parzianello and A

    L. Parzianello and A. Czajka. Saliency-guided textured contact lens-aware iris recognition. InIEEE Winter Conference on Applications of Computer Vision, pages 330–337, 2022. 1

    2022

  28. [36]

    Y . Peng, K. Ma, Y . Zhang, and Z. He. Learning rich features for gait recognition by integrating skeletons and silhouettes.Multimedia Tools and Applications, 83(3):7273–7294, 2024. 8

    2024

  29. [37]

    J. K. Pillai, V . M. Patel, R. Chellappa, and N. K. Ratha. Secure and robust iris recognition using random projections and sparse representations.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 33(9):1877–1893, 2011. 1

    2011

  30. [38]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2, 3, 4, 5, 8, 9

    2021

  31. [39]

    Ranjan, V

    R. Ranjan, V . M. Patel, and R. Chellappa. Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition.IEEE transactions on pattern analysis and machine intelligence, 41(1):121–135, 2017. 1 12

    2017

  32. [40]

    C. Shen, C. Fan, W. Wu, R. Wang, G. Q. Huang, and S. Yu. Lidargait: Benchmarking 3d gait recogni- tion with point clouds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1054–1063, 2023. 2, 3, 5, 6, 7, 8

    2023

  33. [41]

    C. Shen, B. Lin, S. Zhang, X. Yu, G. Q. Huang, and S. Yu. Gait recognition with mask-based regularization. In2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2023. 1, 3

    2023

  34. [42]

    C. Shen, S. Yu, J. Wang, G. Q. Huang, and L. Wang. A comprehensive survey on deep gait recognition: Algorithms, datasets, and challenges.IEEE Transactions on Biometrics, Behavior, and Identity Science,

  35. [43]

    Skean, M

    O. Skean, M. R. Arefin, Y . LeCun, and R. Shwartz-Ziv. Does representation matter? exploring intermediate layers in large language models.arXiv preprint arXiv:2412.09563, 2024. 2, 3

    2024 arXiv

  36. [44]

    Skean, M

    O. Skean, M. R. Arefin, D. Zhao, N. Patel, J. Naghiyev, Y . LeCun, and R. Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models.arXiv preprint arXiv:2502.02013, 2025. 2, 3, 5

    2025 arXiv

  37. [45]

    Y . Su, M. Kim, F. Liu, A. Jain, and X. Liu. Open-set biometrics: Beyond good closed-set models. In European Conference on Computer Vision, pages 243–261. Springer, 2024. 3

    2024

  38. [46]

    Q. Sun, M. Pickett, A. K. Nain, and L. Jones. Transformer layers as painters. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25219–25227, 2025. 2, 3, 6

    2025

  39. [47]

    Teepe, A

    T. Teepe, A. Khan, J. Gilg, F. Herzog, S. Hörmann, and G. Rigoll. Gaitgraph: Graph Convolutional Network for Skeleton-Based Gait Recognition. InInternational Conference on Image Processing (ICIP), pages 2314–2318, 2021. 1, 3

    2021

  40. [48]

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang, et al. Deep high-resolution representation learning for visual recognition.IEEE transactions on pattern analysis and machine intelligence, 43(10):3349–3364, 2020. 9

    2020

  41. [49]

    J. Wang, S. Hou, X. Guo, Y . Huang, Y . Huang, T. Zhang, and L. Wang. Gaitc 3 i: Robust cross-covariate gait recognition via causal intervention.IEEE Transactions on Circuits and Systems for Video Technology,

  42. [50]

    R. Wang, C. Shen, C. Fan, G. Q. Huang, and S. Yu. Pointgait: Boosting end-to-end 3d gait recognition with point clouds via spatiotemporal modeling. In2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2023. 4

  43. [51]

    R. Wang, C. Shen, M. J. Marin-Jimenez, G. Q. Huang, and S. Yu. Cross-modality gait recognition: Bridging lidar and camera modalities for human identification. In2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–11. IEEE, 2024. 4

    2024

  44. [52]

    Z.-Y . Wang, J. Liu, R. P. Kathirvel, C. P. Lau, and R. Chellappa. Hypergait: A video-based multitask net- work for gait recognition and human attribute estimation at range and altitude. In2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–9. IEEE, 2024. 3

    2024

  45. [53]

    S. Yang, J. Wang, S. Hou, X. Liu, C. Cao, L. Wang, and Y . Huang. Bridging gait recognition and large language models sequence modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page in press, 2025. 1, 3

    2025

  46. [54]

    D. Ye, C. Fan, J. Ma, X. Liu, and S. Yu. Biggait: Learning gait representation you want by large vision models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 200–210, 2024. 1, 3, 4, 8, 9

    2024

  47. [55]

    S. Yu, D. Tan, and T. Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In18th International Conference on Pattern Recognition (ICPR’06), volume 4, pages 441–444. IEEE, 2006. 2, 5, 6, 7, 8

    2006

  48. [56]

    Zheng, X

    J. Zheng, X. Liu, W. Liu, L. He, C. Yan, and T. Mei. Gait recognition in the wild with dense 3d representations and a benchmark. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 4

    2022

  49. [57]

    Zheng, X

    J. Zheng, X. Liu, S. Wang, L. Wang, C. Yan, and W. Liu. Parsing is all you need for accurate gait recognition in the wild. InProceedings of the 31st ACM International Conference on Multimedia, pages 116–124, 2023. 1, 3

    2023

  50. [58]

    Zheng, X

    J. Zheng, X. Liu, B. Zhang, C. Yan, J. Zhang, W. Liu, and Y . Zhang. It takes two: Accurate gait recognition in the wild via cross-granularity alignment. InProceedings of the ACM International Conference on Multimedia (ACM MM), pages 8786–8794, 2024. 8 13

    2024

  51. [59]

    S. Zou, C. Fan, J. Xiong, C. Shen, S. Yu, and J. Tang. Cross-Covariate Gait Recognition: A Benchmark. In Association for the Advancement of Artificial Intelligence (AAAI), page [in press], 2024. 1, 2, 5, 6, 7, 8, 9

    2024

  52. [60]

    S. Zou, J. Xiong, C. Fan, C. Shen, S. Yu, and J. Tang. A multi-stage adaptive feature fusion neural network for multimodal gait recognition.IEEE Transactions on Biometrics, Behavior, and Identity Science, 2024. 4 14

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.