Pith. sign in

REVIEW 3 major objections 3 minor 61 references

On Denoising Walking Videos for Gait Recognition

T0 review · 3 major / 3 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 This paper claims that passing walking videos through a pretrained diffusion model and then converting its features into per-pixel direction vectors yields state-of-the-art gait recognition, including mean rank-1 89.5 on CCPG versus 87.2…

desk verdict Useful empirical paper with a novel diffusion-based denoiser for gait, but the SOTA claim rests on test-set-selected hyperparameters and an underspecified diffusion step; deserves peer review with requests for a validation split. read the letter →

arxiv 2505.18582 v1 pith:TUNDXDBV submitted 2025-05-24 cs.CV cs.AI

Dongyang Jin , Chao Fan , Jingzhe Ma , Jingkai Zhou , Weihua Chen , Shiqi Yu This is my paper · ORCID
classification cs.CVcs.AI
keywords gaitrecognitiondenoisingdiffusionStablefeaturesFeatureFieldmatchingcloth-changingcross-domaingeneralizationRGB
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

This paper tries to establish that treating gait recognition as a two-stage denoising problem—first filtering identity-irrelevant cues with a pretrained diffusion model, then condensing the remaining features into direction vectors—produces recognizers that beat prior state of the art on cloth-changing, multi-view RGB gait benchmarks. Gait videos carry both identity and nuisance information such as clothing texture, color, and background, while silhouette and pose inputs remove nuisances but discard too much structure. The authors show that a single forward pass through Stable Diffusion at a carefully chosen timestep partially suppresses the nuisances while retaining shape and motion, and that their Feature Matching module, which converts multi-channel diffusion features into per-pixel two-dimensional direction vectors called static and dynamic Gait Feature Fields, removes residual noise. If the reported gains hold, the work offers a practical path for RGB-based gait recognition under clothing changes and a new use of generative models for a discriminative biometric task.

What carries the argument

Two load-bearing mechanisms carry the argument. The first is the pretrained Latent Diffusion Model, Stable Diffusion, used as a fixed knowledge-driven denoiser: $F_l = \epsilon_\theta(E(I_l), t)$ is a single UNet pass on the clean encoder output at timestep $t$, chosen by a sweep that peaks at $t=700$. The second is the geometry-driven Feature Matching module: query and key features are projected by separate convolutional stacks, masked by silhouettes, and compared over a $(2\Delta h+1) \times (2\Delta w+1)$ neighborhood; the softmax similarity distribution over a fixed direction template $\mathcal{T}$ yields per-pixel two-dimensional direction vectors $G^{\text{Static}}$ (within-frame, $\Delta l=0$) and $G^{\text{Dynamic}}$ (cross-frame, $\Delta l>0$). The output Gait Feature Field is flow-like: the static field traces local appearance gradients and avoids texture-rich regions, while the dynamic field aligns with moving body parts, and random zero-padding of high-magnitude static vectors forces texture-invariant learning.

What would settle it

Re-run the CCPG full cloth-changing experiment with Eq. (3) replaced by an identity mapping on the encoder output, or by a standard noisy latent $z_t = \sqrt{1-\beta_t}E(I_l)+\sqrt{\beta_t}\epsilon$ with the same $t$ and a fixed seed; if rank-1 stays near the reported 84.0, the diffusion-specific one-step denoising is not the cause of the improvement, and if the proper noising step changes accuracy substantially, the paper's stated mechanism is misspecified.

Watch

Extended reading notes

Core claim

The central claim is that diffusion-based knowledge-driven denoising combined with geometry-driven feature vectorization yields a new state of the art for RGB gait recognition. Concretely, DenoisingGait takes each video frame, encodes it with Stable Diffusion's image encoder, and applies one UNet denoising step at timestep t=700, the empirically optimal granularity, then masks the background with silhouettes and runs within-frame and cross-frame Feature Matching. The matching step computes softmax similarities between each query pixel and its local neighbors, multiplies by a fixed direction template, and outputs two-channel static and dynamic Gait Feature Fields; a texture-suppression operation zeroes high-magnitude static vectors during training. Experiments on CCPG, CASIA-B*, and SUSTech1K show improvements over prior state of the art in most within- and cross-domain settings, with the main gains on full cloth-changing sequences (CCPG CL +1.4 over the prior best RGB method) and a +2.7 overall cross-domain gain on SUSTech1K.

Load-bearing premise

The paper relies on Eq. (3) treating $\epsilon_\theta(E(I_l), t)$ as a one-step denoising operation, but the standard latent diffusion reverse process applies to a noisy latent $z_t$, and the paper never specifies how $z_t$ is obtained from the clean encoder output; if this operation is not a true denoising step, the mechanism behind the gains is different from what is claimed.

Editorial extensions

If this is right

  • A moderate diffusion timestep filters clothing texture and color while preserving body shape; the reported 5.3-point jump over no-denoising on CCPG full cloth-changing suggests timestep choice is the main knowledge-driven tuning knob.
  • Silhouette masking plus within-frame matching contributes most of the geometry-driven gain; cross-frame matching alone underperforms static features, so motion fields supplement rather than replace appearance.
  • Texture suppression yields +3.0 on CCPG CL, implying that explicit suppression of high-texture static directions is what makes the representation cloth-robust.
  • Cross-domain results improve by +2.7 overall on SUSTech1K when training on CCPG, indicating the learned fields generalize better than prior RGB methods to unseen domains.

Reading between the lines

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

  • Beyond the paper: because Eq. (3) feeds a clean latent into a UNet trained on noisy latents, the 'denoising' may act as a learned nonlinear smoother rather than a true reverse diffusion step; if so, the reported gains could transfer to other fixed feature extractors with comparable smoothing, which is testable by swapping in a frozen convolutional autoencoder.
  • Beyond the paper: the per-pixel direction-vector representation does not need to be tied to gait; the same silhouette-masked within- and cross-frame matching could be applied to cloth-changing person re-identification or video-based action recognition, where texture invariance and local motion are both discriminative.
  • Beyond the paper: the night-scene weakness attributed to low-quality silhouettes suggests that replacing binary masks with soft, learned human segmentation could extend the method to low-light domains without changing the rest of the pipeline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes DenoisingGait, a gait recognition method that combines knowledge-driven denoising from a pretrained Stable Diffusion model with a geometry-driven Feature Matching module. The diffusion features are fed into a matching module that produces static and dynamic Gait Feature Fields, and a texture suppression operation is applied during training. The authors report state-of-the-art results on CCPG, CASIA-B*, and SUSTech1K in both within- and cross-domain evaluations, with code released.

Significance. If validated, the work would introduce a novel and interesting use of diffusion models for gait recognition, along with a new flow-like representation that demonstrably suppresses clothing texture in ablations. The paper includes component-wise ablations (Tables 5-7) and cross-domain experiments, and the code is publicly available. However, the empirical claims are currently not fully supported because key hyperparameters appear to be selected on the test set and no error bars are reported. The central SOTA result therefore needs additional verification before the contribution can be fully assessed.

major comments (3)
  1. [Sec. 3.1, Eq. (3)] In Eq. (3), the paper defines Fl = εθ(E(Il), t) as 'one-step denoising' using the pretrained Stable Diffusion model. Under the LDM formalism of Eqs. (1)-(2), the UNet εθ is trained to predict the noise from a noisy latent zt, not from the clean encoder output E(Il). The forward noising step that produces zt is never described in Sec. 3.1. As written, the operation is not the standard reverse-process denoising step, so the interpretation of Fl as denoised diffusion features is not grounded in the diffusion formulation. Please specify how zt is computed from E(Il) (e.g., by applying Eq. (1) with a particular β_t), or, if εθ is applied to a clean latent, justify this out-of-distribution usage and re-evaluate the reported benefits accordingly.
  2. [Sec. 4, Fig. 2(b), Tables 6-7] The timestep t=700 is selected by evaluating a baseline on the CCPG full cloth-changing case (Fig. 2b), and the texture-suppression threshold m and zero-padding probability p are selected using ablations on CCPG (Tables 6-7). No validation split is described in Sec. 4, so these hyperparameters appear to be chosen on the same CCPG test protocol later used for the SOTA numbers in Table 1. Since Table 7 shows that texture suppression alone improves CL accuracy by +3.0% (81.0→84.0) and Table 1 shows gains over prior SOTA of only +0.1 (CL), +2.1 (UP), +2.9 (DN), and +2.8 (BG), test-set selection could materially alter the comparison. Please introduce a validation split and report results with hyperparameters chosen on that split, or otherwise demonstrate the robustness of the reported SOTA to the choice of t, m, and p.
  3. [Sec. 4.3, Tables 1, 3, 4] All presented accuracies are single runs without error bars or multiple seeds. Given that several improvements over prior methods are within 1-3 percentage points (e.g., +0.1 in CL on CCPG in Table 1) and the hyperparameters are tuned on the same benchmark, the central SOTA claim is not statistically established. Please report the mean and standard deviation over at least three independent runs, or otherwise provide a statistical comparison with the baselines.
minor comments (3)
  1. [Sec. 4.4] The paragraph after Table 6 contains a typo: 'gait feature fiedl' should be 'gait feature field'.
  2. [Fig. 2(b)] The caption does not define the red line; please state explicitly that it represents the baseline without the denoising operation εθ in Eq. (3).
  3. [Sec. 4.1] The statement that all implementations strictly adhere to the protocols set by dataset publishers is not accompanied by a description of the CCPG gait vs. ReID evaluation protocols used in Table 1; please clarify which protocol is used for the main results.

Circularity Check

2 steps flagged · score 3.0 of 10

No equation-level circularity, but the central CCPG SOTA claim is partly weakened because key hyperparameters t and m are selected on the same CCPG evaluation protocol later used to report the final SOTA numbers.

  1. fitted input called prediction [Sec. 3.1, Fig. 2(b); used in Sec. 4.3, Table 1]
    "When the timestep t is gradually decreased from T = 1, 000 to 100 with an interval of 200, as shown in Figure 2 (b), the rank-1 accuracy of our baseline initially rises and then falls, and the peak is at t = 700."

    The diffusion timestep t is chosen by sweeping the baseline's rank-1 accuracy on the CCPG full cloth-changing case, which is the same CCPG gait evaluation protocol later used to report the SOTA numbers in Table 1. No validation split is described in Sec. 4, so t=700 is effectively fitted to the test protocol. The final CCPG result for DenoisingGait therefore incorporates a hyperparameter already optimized on that protocol, so the headline 'new SoTA' claim is not an independent prediction of the method with a pre-specified t.

  2. fitted input called prediction [Sec. 4.2 item (3) and Sec. 4.4, Table 7]
    "(3) (H, W, d) = (768, 384, 8), (∆h, ∆w) = (3, 3), m=0.5 for Section 3;"

    The texture-suppression threshold m (and the probability p in Eq. 9) is not set on an independent validation split. The paper states 'All ablation experiments are conducted on CCPG' and reports that enabling texture suppression improves CL Rank-1 by +3.0%. Choosing m by CCPG Rank-1 accuracy and then reporting the CCPG Rank-1 of the chosen configuration as SOTA is a fitted-input-called-prediction loop: the +3.0% gain in Table 7 is the same selection signal used to fix the hyperparameter, not an out-of-sample verification of the final claim.

full rationale

I walked the paper's derivation chain. There is no equation-level circularity: the Feature Matching module is trained with identity labels, and no formula in the paper defines the SOTA result in terms of its own inputs. The diffusion features are extracted from a pretrained Stable Diffusion model, and the geometry-driven module is a separate learnable component, so the method is not a renaming of a known result. The self-citations (GaitBase, OpenGait, SkeletonGait++, etc.) are used as backbones and baselines, not as a load-bearing uniqueness theorem, so they do not raise the circularity score. The main concern is empirical: t=700 and m=0.5 are selected using the same CCPG protocol on which the final SOTA table is reported, with no described validation split, and the paper itself reports that texture suppression alone changes CL accuracy by +3.0%. That is a fitted-input-called-prediction issue rather than a construction-level circularity. Separately, Eq. 3 passes the clean encoder output E(Il) to epsilon_theta where the LDM formulation in Eqs. 1-2 expects a noised latent z_t; this is a mechanistic validity concern but not circularity. Overall score 3 reflects the test-protocol hyperparameter selection weakening the central CCPG claim, while the method retains independent content and honest non-circular components.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on pretrained Stable Diffusion as a denoiser, a hand-selected timestep, a texture-suppression threshold, and silhouette masks. The learned Feature Matching module is trained with identity labels, so it is not a free parameter in the ledger, but its design choices such as window size and probability p are under-specified.

free parameters (3)
  • Diffusion timestep t = 700
    Chosen from a rank-1 accuracy scan on the CCPG full cloth-changing protocol (Figure 2b) and then fixed for all experiments.
  • Texture suppression threshold m = 0.5
    Given in Sec. 4.2 without derivation; controls whether a pixel is zero-padded in Eq. 9.
  • Zero-padding probability p = not reported
    Introduced in Eq. 9; no value is given in the implementation details, so the texture suppression operation is not fully specified.
assumptions (6)
  • domain assumption Latent Diffusion formulation (Eqs. 1-2) and pretrained Stable Diffusion v1.5 can provide useful gait features.
    The method relies on a frozen SD 1.5 model as a knowledge source without fine-tuning (Sec. 3.1).
  • ad hoc to paper One-step denoising with a clean encoder output is a valid operation.
    Eq. 3 omits the forward noising step; either the UNet is fed z0 directly, which is inconsistent with LDM inference, or an undocumented noising step exists.
  • ad hoc to paper Timestep t=700 generalizes across datasets and test protocols.
    t=700 is selected on the CCPG full cloth-changing protocol (Figure 2b) and then fixed for all datasets; no validation split is used.
  • domain assumption The magnitude of the static Gait Feature Field reflects texture intensity, so zeroing high-magnitude pixels removes texture cues.
    Stated as an observation in Sec. 3.3 and implemented as Eq. 9, but no quantitative verification or causal analysis is provided.
  • domain assumption Silhouette masks are accurate enough for background removal in all test conditions.
    Features are masked by silhouettes in Eq. 4; the paper acknowledges night-time silhouettes are low quality and NT accuracy is limited (Sec. 4.3).
  • standard math Standard diffusion math (Markov noising, simplified LDM objective) is correct.
    The paper uses Eqs. 1-2 from Ho et al. and Rombach et al. as background; no issue with the math itself, only with its application in Eq. 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Denoising Walking Videos for Gait Recognition." pith.science (2026). https://pith.science/paper/TUNDXDBV

@misc{pith2026250518582,
  author       = {Pith},
  title        = {Pith review of: On Denoising Walking Videos for Gait Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TUNDXDBV}},
  note         = {Machine review of arXiv:2505.18582}
}
read the original abstract

To capture individual gait patterns, excluding identity-irrelevant cues in walking videos, such as clothing texture and color, remains a persistent challenge for vision-based gait recognition. Traditional silhouette- and pose-based methods, though theoretically effective at removing such distractions, often fall short of high accuracy due to their sparse and less informative inputs. Emerging end-to-end methods address this by directly denoising RGB videos using human priors. Building on this trend, we propose DenoisingGait, a novel gait denoising method. Inspired by the philosophy that "what I cannot create, I do not understand", we turn to generative diffusion models, uncovering how they partially filter out irrelevant factors for gait understanding. Additionally, we introduce a geometry-driven Feature Matching module, which, combined with background removal via human silhouettes, condenses the multi-channel diffusion features at each foreground pixel into a two-channel direction vector. Specifically, the proposed within- and cross-frame matching respectively capture the local vectorized structures of gait appearance and motion, producing a novel flow-like gait representation termed Gait Feature Field, which further reduces residual noise in diffusion features. Experiments on the CCPG, CASIA-B*, and SUSTech1K datasets demonstrate that DenoisingGait achieves a new SoTA performance in most cases for both within- and cross-domain evaluations. Code is available at https://github.com/ShiqiYu/OpenGait.

Figures

Figures reproduced from arXiv: 2505.18582 by the authors.

Figure 1
Figure 1. The proposed knowledge-driven denoising, derived from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) A simple baseline on diffusion models for gait representation learning. (b) The rank-1 accuracy of our baseline with varying [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) our Feature Matching module. (b) Assignment of Direction (AoD). Background removal operation uses pre-extracted [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Raw RGB images. (b) Static gait feature field, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 47 canonical work pages

  1. [1]

    Segdiff: Image segmentation with diffusion proba- bilistic models

    Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion proba- bilistic models. arXiv preprint arXiv:2112.00390, 2021. 2

    arXiv 2021

  2. [2]

    Gaitset: Cross-view gait recognition through utilizing gait as a deep set

    Hanqing Chao, Kun Wang, Yiwei He, Junping Zhang, and Jianfeng Feng. Gaitset: Cross-view gait recognition through utilizing gait as a deep set. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3467–3478, 2022. 1, 2, 6, 7

    work page 2022

  3. [3]

    Dif- fusiondet: Diffusion model for object detection

    Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Dif- fusiondet: Diffusion model for object detection. In Proceed- ings of the IEEE/CVF international conference on computer vision, pages 19830–19843, 2023. 2

    2023

  4. [4]

    Perception pri- oritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472–11481, 2022. 2

    work page 2022

  5. [5]

    Text-to-image diffusion mod- els are zero shot classifiers

    Kevin Clark and Priyank Jaini. Text-to-image diffusion mod- els are zero shot classifiers. Advances in Neural Information Processing Systems, 36, 2024. 2

    work page 2024

  6. [6]

    On analyzing generative and denoising capabil- ities of diffusion-based deep generative models

    Kamil Deja, Anna Kuzina, Tomasz Trzcinski, and Jakub Tomczak. On analyzing generative and denoising capabil- ities of diffusion-based deep generative models. Advances in Neural Information Processing Systems, 35:26218–26229,

    work page

  7. [7]

    GaitPart: Temporal part-based model for gait recognition

    Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. GaitPart: Temporal part-based model for gait recognition. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 14225–14233, 2020. 1, 2, 5, 6, 7

    work page 2020

  8. [8]

    Ex- ploring deep models for practical gait recognition

    Chao Fan, Saihui Hou, Yongzhen Huang, and Shiqi Yu. Ex- ploring deep models for practical gait recognition. arXiv preprint arXiv:2303.03301, 2023. 2, 6, 7

    arXiv 2023

Show all 61 references
  1. [9]

    Opengait: Revisiting gait recognition towards better practicality

    Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, and Shiqi Yu. Opengait: Revisiting gait recognition towards better practicality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9707–9716, 2023. 4, 5, 6, 7, 8

    2023

  2. [10]

    Open- gait: A comprehensive benchmark study for gait recognition towards better practicality.arXiv preprint arXiv:2405.09138,

    Chao Fan, Saihui Hou, Junhao Liang, Chuanfu Shen, Jingzhe Ma, Dongyang Jin, Yongzhen Huang, and Shiqi Yu. Open- gait: A comprehensive benchmark study for gait recognition towards better practicality.arXiv preprint arXiv:2405.09138,

    arXiv

  3. [11]

    Skeletongait: Gait recognition using skeleton maps

    Chao Fan, Jingzhe Ma, Dongyang Jin, Chuanfu Shen, and Shiqi Yu. Skeletongait: Gait recognition using skeleton maps. In Proceedings of the AAAI conference on artificial intelligence, pages 1662–1669, 2024. 1, 2, 5, 6, 7

    2024

  4. [12]

    Data augmentation for object detec- tion via controllable diffusion models

    Haoyang Fang, Boran Han, Shuai Zhang, Su Zhou, Cuixiong Hu, and Wen-Ming Ye. Data augmentation for object detec- tion via controllable diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1257–1266, 2024. 2

    2024

  5. [13]

    Gaitfusion: Explor- ing the fusion of silhouettes and optical flow for gait recog- nition

    Yuxiang Feng, Jiabin Yuan, and Lili Fan. Gaitfusion: Explor- ing the fusion of silhouettes and optical flow for gait recog- nition. In Artificial Neural Networks and Machine Learning – ICANN 2023, pages 88–99, Cham, 2023. Springer Nature Switzerland. 3

    2023

  6. [14]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2, 3

    2020

  7. [15]

    Ex- ploring more from multiple gait modalities for human iden- tification

    Dongyang Jin, Chao Fan, Weihua Chen, and Shiqi Yu. Ex- ploring more from multiple gait modalities for human iden- tification. arXiv preprint arXiv:2412.11495, 2024. 6, 7

    2024 arXiv

  8. [16]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9492...

    2024

  9. [17]

    An in-depth ex- ploration of person re-identification and gait recognition in cloth-changing conditions

    Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth ex- ploration of person re-identification and gait recognition in cloth-changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2023

  10. [18]

    End-to-end model-based gait recog- nition

    Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recog- nition. In Proceedings of the Asian Conference on Computer Vision, 2020. 1, 2

    2020

  11. [19]

    Gaitedge: Beyond plain end-to-end gait recognition for better practicality

    Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In Com- puter Vision – ECCV 2022, 2022. 2, 6, 7

    2022

  12. [20]

    Pose-based temporal-spatial network (ptsn) for gait recognition with carrying and clothing varia- tions

    Rijun Liao, Chunshui Cao, Edel B Garcia, Shiqi Yu, and Yongzhen Huang. Pose-based temporal-spatial network (ptsn) for gait recognition with carrying and clothing varia- tions. In Chinese conference on biometric recognition, pages 474–483. Springer, 2017. 1, 2

    2017

  13. [21]

    Gait recognition via effective global-local feature representation and local tem- poral aggregation

    Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local tem- poral aggregation. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 14648–14656,

  14. [22]

    Zongyi Liu and S. Sarkar. Improved gait recognition by gait dynamics normalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(6):863–876, 2006. 3

    2006

  15. [23]

    Distinctive image features from scale- invariant keypoints

    David G Lowe. Distinctive image features from scale- invariant keypoints. International journal of computer vi- sion, 60:91–110, 2004. 2

    2004

  16. [24]

    Pedes- trian attribute editing for gait recognition and anonymization

    Jingzhe Ma, Dingqiang Ye, Chao Fan, and Shiqi Yu. Pedes- trian attribute editing for gait recognition and anonymization. arXiv preprint arXiv:2303.05076, 2023. 5

    2023 arXiv

  17. [25]

    Automatic recognition by gait

    Mark S Nixon and John N Carter. Automatic recognition by gait. Proceedings of the IEEE, 94(11):2013–2024, 2006. 1

    2013

  18. [26]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 8

    2023 arXiv

  19. [27]

    Learning rich features for gait recognition by integrating skeletons and silhouettes

    Yunjie Peng, Kang Ma, Yang Zhang, and Zhiqiang He. Learning rich features for gait recognition by integrating skeletons and silhouettes. Multimedia Tools and Applica- tions, 83(3):7273–7294, 2024. 6, 7

    2024

  20. [28]

    Tiger: Time-varying denoising model for 3d point cloud generation with diffusion process

    Zhiyuan Ren, Minchul Kim, Feng Liu, and Xiaoming Liu. Tiger: Time-varying denoising model for 3d point cloud generation with diffusion process. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9462–9471, 2024. 2, 4

    2024

  21. [29]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 6, 8

    2022

  22. [30]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Inter- vention – MICCAI 2015, pages 234–241, 2015. 3

    2015

  23. [31]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE in- ternational conference on computer vision , pages 618–626,

  24. [32]

    Huang, and Shiqi Yu

    Chuanfu Shen, Chao Fan, Wei Wu, Rui Wang, George Q. Huang, and Shiqi Yu. Lidargait: Benchmarking 3d gait recognition with point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1054–1063, 2023. 2, 6, 7

    2023

  25. [33]

    Lidargait: Benchmarking 3d gait recognition with point clouds

    Chuanfu Shen, Chao Fan, Wei Wu, Rui Wang, George Q Huang, and Shiqi Yu. Lidargait: Benchmarking 3d gait recognition with point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1054–1063, 2023. 4

    2023

  26. [34]

    Gait recognition with mask-based reg- ularization

    Chuanfu Shen, Beibei Lin, Shunli Zhang, Xin Yu, George Q Huang, and Shiqi Yu. Gait recognition with mask-based reg- ularization. In 2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2023. 1

    2023

  27. [35]

    A comprehensive survey on deep gait recogni- tion: Algorithms, datasets, and challenges

    Chuanfu Shen, Shiqi Yu, Jilong Wang, George Q Huang, and Liang Wang. A comprehensive survey on deep gait recogni- tion: Algorithms, datasets, and challenges. IEEE Transac- tions on Biometrics, Behavior, and Identity Science , 2024. 1

    2024

  28. [36]

    Videoflow: Exploiting temporal cues for multi-frame optical flow estimation

    Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Com...

    2023

  29. [37]

    Gaitnet: An end-to-end network for gait based human identification

    Chunfeng Song, Yongzhen Huang, Yan Huang, Ning Jia, and Liang Wang. Gaitnet: An end-to-end network for gait based human identification. Pattern recognition, 96:106988, 2019. 8

    2019

  30. [38]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In International Conference on Learning Representations, 2021. 2

    2021

  31. [39]

    High-resolution image re- construction with latent diffusion models from human brain activity

    Yu Takagi and Shinji Nishimoto. High-resolution image re- construction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14453– 14463, 2023. 2

    2023

  32. [40]

    General tensor discriminant analysis and gabor fea- tures for gait recognition.IEEE transactions on pattern anal- ysis and machine intelligence, 29(10):1700–1715, 2007

    Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J May- bank. General tensor discriminant analysis and gabor fea- tures for gait recognition.IEEE transactions on pattern anal- ysis and machine intelligence, 29(10):1700–1715, 2007. 3

    2007

  33. [41]

    Gaitgraph: graph convo- lutional network for skeleton-based gait recognition

    Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Ste- fan H¨ormann, and Gerhard Rigoll. Gaitgraph: graph convo- lutional network for skeleton-based gait recognition. In2021 IEEE International Conference on Image Processing (ICIP), pages 2314–2318. IEEE, 2021. 1, 2

    2021

  34. [42]

    Diffuse attend and segment: Un- supervised zero-shot segmentation using stable diffusion

    Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse attend and segment: Un- supervised zero-shot segmentation using stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3554–3563...

    2024

  35. [43]

    Pointgait: Boosting end-to-end 3d gait recogni- tion with point clouds via spatiotemporal modeling

    Rui Wang, Chuanfu Shen, Chao Fan, George Q Huang, and Shiqi Yu. Pointgait: Boosting end-to-end 3d gait recogni- tion with point clouds via spatiotemporal modeling. In 2023 IEEE International Joint Conference on Biometrics (IJCB) , pages 1–10. IEEE, 2023. 5

    2023

  36. [44]

    Cross-modality gait recog- nition: Bridging lidar and camera modalities for human iden- tification

    Rui Wang, Chuanfu Shen, Manuel J Marin-Jimenez, George Q Huang, and Shiqi Yu. Cross-modality gait recog- nition: Bridging lidar and camera modalities for human iden- tification. In 2024 IEEE International Joint Conference on Biometrics (IJCB), pages 1–11. IEEE, 2024. 5

    2024

  37. [45]

    Gmflow: Learning optical flow via global matching

    Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121–8130,

  38. [46]

    Unifying flow, stereo and depth estimation

    Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6

    2023

  39. [47]

    Attention-based gait recognition network with novel partial representation pgofi based on prior motion information

    Jian Xu, Hai Li, and Shujuan Hou. Attention-based gait recognition network with novel partial representation pgofi based on prior motion information. Digital Signal Process- ing, 133:103845, 2023. 3

    2023

  40. [48]

    Open-vocabulary panop- tic segmentation with text-to-image diffusion models

    Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiao- long Wang, and Shalini De Mello. Open-vocabulary panop- tic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2955–2966, 2023. 2

    2023

  41. [49]

    Diffusion model as repre- sentation learner

    Xingyi Yang and Xinchao Wang. Diffusion model as repre- sentation learner. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 18938–18949,

  42. [50]

    Biggait: Learning gait representation you want by large vision models

    Dingqiang Ye, Chao Fan, Jingzhe Ma, Xiaoming Liu, and Shiqi Yu. Biggait: Learning gait representation you want by large vision models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6, 7, 8

    2024

  43. [51]

    Gait recognition based on gait optical flow network with inherent feature pyramid

    Hongyi Ye, Tanfeng Sun, and Ke Xu. Gait recognition based on gait optical flow network with inherent feature pyramid. Applied Sciences, 13(19):10975, 2023. 3

    2023

  44. [52]

    A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition

    Shiqi Yu, Daoliang Tan, and Tieniu Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In 18th International Con- ference on Pattern Recognition (ICPR’06) , pages 441–444. IEEE, 2006. 2, 6, 7

    2006

  45. [53]

    Representation alignment for generation: Training diffu- sion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffu- sion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 2

    arXiv

  46. [54]

    Exploring diffusion time- steps for unsupervised representation learning

    Zhongqi Yue, Jiankun Wang, Qianru Sun, Lei Ji, Eric I Chang, Hanwang Zhang, et al. Exploring diffusion time- steps for unsupervised representation learning. arXiv preprint arXiv:2401.11430, 2024. 2

    2024 arXiv

  47. [55]

    On learning disentangled representations for gait recogni- tion

    Ziyuan Zhang, Luan Tran, Feng Liu, and Xiaoming Liu. On learning disentangled representations for gait recogni- tion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2, 8

    2020

  48. [56]

    Gait recognition in the wild with dense 3d representations and a benchmark

    Jinkai Zheng, Xinchen Liu, Wu Liu, Lingxiao He, Cheng- gang Yan, and Tao Mei. Gait recognition in the wild with dense 3d representations and a benchmark. In IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

  49. [57]

    Parsing is all you need for accurate gait recognition in the wild

    Jinkai Zheng, Xinchen Liu, Shuai Wang, Lihao Wang, Chenggang Yan, and Wu Liu. Parsing is all you need for accurate gait recognition in the wild. In Proceedings of the 31st ACM International Conference on Multimedia , pages 116–124, 2023. 1, 2

    2023

  50. [58]

    It takes two: Accurate gait recognition in the wild via cross-granularity alignment

    Jinkai Zheng, Xinchen Liu, Boyue Zhang, Chenggang Yan, Jiyong Zhang, Wu Liu, and Yongdong Zhang. It takes two: Accurate gait recognition in the wild via cross-granularity alignment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8786–8794, 2024. 6

    2024

  51. [59]

    Dpmesh: Exploiting diffusion prior for occluded human mesh recovery

    Yixuan Zhu, Ao Li, Yansong Tang, Wenliang Zhao, Jie Zhou, and Jiwen Lu. Dpmesh: Exploiting diffusion prior for occluded human mesh recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1101–1110, 2024. 2, 4

    2024

  52. [60]

    Cross-covariate gait recognition: A bench- mark

    Shinan Zou, Chao Fan, Jianbo Xiong, Chuanfu Shen, Shiqi Yu, and Jin Tang. Cross-covariate gait recognition: A bench- mark. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7855–7863, 2024. 1, 2

    2024

  53. [61]

    A multi-stage adaptive feature fusion neu- ral network for multimodal gait recognition

    Shinan Zou, Jianbo Xiong, Chao Fan, Chuanfu Shen, Shiqi Yu, and Jin Tang. A multi-stage adaptive feature fusion neu- ral network for multimodal gait recognition. IEEE Trans- actions on Biometrics, Behavior, and Identity Science, 2024. 5

    2024

Pith tools

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