Learning A Unified Template for Gait Recognition

Panjian Huang; Saihui Hou; Junzhou Huang; Yongzhen Huang

cs.CVSep 16, 2026

Learning A Unified Template for Gait Recognition

Authors: Panjian Huang, Saihui Hou, Junzhou Huang, Yongzhen Huang

Organizations: School of Artificial Intelligence, Beijing Normal University · Department of Computer Science and Engineering, The University of Texas at Arlington

Abstract

"What I cannot create, I do not understand."Human wisdom reveals that creation is one of the highest forms of learning. For example, Diffusion Models have demonstrated remarkable semantic structure and memory in image generation, understanding, and restoration, which intuitively benefits representation learning. However, current gait networks rarely embrace this perspective, relying primarily on learning by contrasting gait samples under varying complex conditions, leading to semantic inconsistency and uniformity issues. To address these issues, we propose Origins with generative capabilities whose underlying philosophy is that different entities are generated from a unified template, inherently regularizing gait representations within a consistent and diverse semantic space to capture accurate gait differences. Admittedly, learning this unified template is exceedingly challenging, as it requires the comprehensiveness of the template to encompass gait representations with various conditions. Inspired by Diffusion Models, Origins diffuses the unified template into timestep templates for gait generative learning, and meanwhile transfers the unified template for gait representation learning. Especially, gait generative and representation learning serve as a unified framework for end-to-end joint training. Extensive experiments on CASIA-B, CCPG,SUSTech1K, Gait3D, GREW and CCGR-MINI demonstrate that Origins performs unified generative and representation learning, achieving superior performance.

Figures & tables

Figure 1: Origins performs unified generative and representation learning. The unified template $\mathcal{T}$ diffuses to the timestep templates, which generates gait representations. Meanwhile, the unified template $\mathcal{T}$ transfers to capture accurate gait differences.

Figure 2: The unified template is the consistent and diverse space. The different semantic dimensions in this unified template linearly construct each gait representation.

Figure 3: More timestep templates enable better convergence and more precise recognition performance.

Figure 4: The heatmaps show that the unified template generates the missing human body parts for information supplementation and matching ( e.g . , the regions in the red circle of the left and right figures), and rectifies temporal information ( e.g . , the region in the red circle of the middle figure).

Figure 5: The gait network with Origins. HP represents Horizontal Partition, and Visual Stage consists of basic convolution blocks. After the gait silhouette sequence passes through Visual Encoder and HP, the part representations are fed into the respective Origins and Recognition Head for generative and representation learning.

Figure 6: The overview of Origins, best viewed in colors and shapes. The $\mathcal{N}$ , $\mathcal{C}_{4}$ and $\mathcal{Z}$ denote the number, dimension and multi-heads of timestep templates. The unified template $\mathcal{T}$ diffuses timestep templates by time serialization for generation and transfers information to gait representations for recognition.

Environment	Dataset	Train		Test		Condition	Stage	Channels	Strides
Environment	Dataset	Id.	Seq.	Id.	Seq.	Condition	[ $\mathcal{D}_{1}$ , $\mathcal{D}_{2}$ , $\mathcal{D}_{3}$ , $\mathcal{D}_{4}$ ]	[ $\mathcal{C}_{1}$ , $\mathcal{C}_{2}$ , $\mathcal{C}_{3}$ , $\mathcal{C}_{4}$ ]	Strides
Constrained	CASIA-B [ 66 ]	74	8,140	50	5,500	CV, BG, CL	[1, 1, 1. -]	[64, 128, 256, -]	[1, 2, 1, -]
Constrained	CCPG [ 29 ]	100	8,187	100	8,095	CV, BG, CL	[1, 1, 1, 1]	[64, 128, 256, 512]	[1, 2, 2, 1]
In-the-wild	SUSTech1K [ 45 ]	200	5988	850	19,228	Real-world	[1, 1, 1, 1]	[64, 128, 256, 512]	[1, 2, 2, 1]
	CCGR-MINI [ 73 ]	571	27507	399	20377	Real-world	[1, 4, 4, 1]	[64, 128, 256, 512]	[1, 2, 2, 1]
	Gait3D [ 70 ]	3,000	18,940	1,000	6,369	Real-world	[1, 4, 4, 1]	[64, 128, 256, 512]	[1, 2, 2, 1]

Table 1: DataBase and Architectures. Id. and Seq. denote the number of identities and sequences. CV, BG and CL refer to cross-view and carrying bags and cross-clothing conditions.

\mathcal{D}

and

\mathcal{C}

denote the number of conv blocks and the channels in each visual stage.

Paradigm	Method	Venue	Gait Evaluation Protocol					ReID Evaluation Protocol
Paradigm	Method	Venue	CL	UP	DN	BG	Mean	CL	UP	DN	BG	Mean
Model	GaitGraph2 [ 48 ]	CVPRW22	5.0	5.3	5.8	6.2	5.1	5.0	5.7	7.3	8.8	6.7
	Gait-TR [ 68 ]	ES23	15.7	18.3	18.5	17.5	17.5	24.3	28.7	31.1	28.1	28.1
	MSGG [ 40 ]	MTA23	29.0	34.5	37.1	33.3	33.5	43.1	52.9	57.4	49.9	50.8
	SkeletonGait [ 11 ]	AAAI24	40.4	48.5	53.0	61.7	50.9	52.4	65.4	72.8	80.9	67.9
Appearance	GaitSet [ 3 ]	AAAI19	60.2	65.2	65.1	68.5	64.8	77.5	85.0	82.9	87.5	83.2

Table 2: The performance comparisons on CCPG under various conditions, including full-body cloth changes (CL), upper-body cloth changes (UP), lower-body cloth changes (DN), and backpacks changes (BG), reported with Rank-1 accuracy (%).

Paradigm	Method	Venue	Probe Sequence (Rank-1)								Overall
Paradigm	Method	Venue	Normal	Bag	Clothing	Carrying	Umbrella	Uniform	Occlusion	Night	Rank-1	Rank-5
Model	GaitGraph2 [ 48 ]	CVPRW22	22.2	18.2	6.8	18.6	13.4	19.2	27.3	16.4	18.6	40.2
	Gait-TR [ 68 ]	ES23	33.3	31.5	21.0	30.4	22.7	34.6	44.9	23.5	30.8	56.0
	MSGG [ 40 ]	MTA23	67.11	66.16	35.92	63.31	61.58	58.07	66.59	17.88	33.8	-
	SkeletonGait [ 11 ]	AAAI24	67.9	63.5	36.5	61.6	58.1	67.2	79.1	50.1	63.0	83.5
Appearance	GaitSet [ 3 ]	AAAI19	69.1	68.2	37.4	65.0	63.1	61.0	67.2	23.0	65.0	84.8

Table 3: The performance comparisons on SUSTech1K are reported with Rank-1 and Rank-5 accuracy (%).

Method	Venue	NM	BG	CL	Mean
GaitSet [ 3 ]	AAAI19	95.0	87.2	70.4	84.2
GaitPart [ 12 ]	CVPR20	96.2	91.5	78.7	88.8
GLN [ 21 ]	ECCV20	96.9	94.0	77.5	89.5
GaitGL [ 32 ]	ICCV21	97.4	94.5	83.6	91.8
QAGait [ 58 ]	AAAI24	97.9	94.6	78.2	90.2
GaitBase [ 10 ]	CVPR23	97.6	94.0	77.4	89.8

Table 4: The performance comparisons on CASIA-B are reported with Rank-1 accuracy (%).

Method	Venue	Rank-1	Rank-5	mAP
GaitSet [ 3 ]	AAAI19	36.7	58.3	30.0
GaitPart [ 12 ]	CVPR20	28.2	47.6	47.6
GaitGL [ 32 ]	ICCV21	29.7	48.5	22.3
SMPLGait [ 70 ]	CVPR22	46.3	64.5	37.2
MTSGait [ 69 ]	MM22	48.7	67.1	37.6
QAGait [ 58 ]	AAAI24	67.0	81.5	56.5

Table 5: The performance comparisons on Gait3D are reported with Rank-1, Rank-5 accuracy and mAP (%).

Method	Venue	Rank-1	Rank-5	Rank-10
GaitSet [ 3 ]	AAAI19	46.3	63.6	70.3
GaitPart [ 12 ]	CVPR20	44.0	60.7	67.3
GaitGL [ 32 ]	ICCV21	47.3	63.6	-
MTSGait [ 69 ]	MM22	55.3	71.3	76.9
QAGait [ 58 ]	AAAI24	59.1	74.0	79.2
GaitBase [ 10 ]	CVPR23	60.1	-	-

Table 6: The performance comparisons on GREW are reported with Rank-1, Rank-5 and Rank-10 accuracy (%).

Method	Venue	Rank-1	mAP	mINP
GaitSet [ 3 ]	AAAI19	13.77	15.39	5.75
GaitPart [ 12 ]	CVPR20	8.02	10.12	3.52
GaitGL [ 32 ]	ICCV21	17.51	18.12	6.85
GaitBase [ 10 ]	CVPR23	26.99	24.89	9.72
DeepGaitV2 [ 9 ]	TPAMI25	39.37	36.01	16.77
Origins-M ( ours )	-	41.45	38.31	24.71

Table 7: The performance comparisons on CCGR-MINI are reported with Rank-1 accuracy, mAP and mINP (%).

Panel of Figure 7 — Figure 7: The “elements” axe is $\mathcal{Z}$ that is the number of multi-heads in the template. The visualization shows that different subjects exhibit differences in the template.

The analysis on the core modules
Method	Gait3D
Method	Rank-1	Rank-5	mAP
Origins-M	75.8	86.8	67.0
-w/o Origins	73.6	86.9	65.1
- w/o Transferring Unified Template	74.4	87.9	67.1
- w/o Diffusing Timestep Templates	75.3	86.4	66.3

Table 8: The core module analysis on Gait3D.

Method	Gait3D			CASIA-B
Method	Rank-1	Rank-5	mAP	NM	BG	CL	Mean
Origins-M	75.8	86.8	67.0	99.3	97.4	90.3	95.7
Baseline	73.6	86.9	65.1	98.2	95.8	84.4	92.8
The analysis on the number of timestep templates
$\mathcal{N}=8$	75.3	87.0	66.5	99.1	97.3	90.4	95.6
$\mathcal{N}=16$	75.8	86.8	67.0	99.3	97.4	90.3	95.7

Table 9: The timestep template analysis on Gait3D and CASIA-B.

Gait Recognition

Gait recognition identifies people from their walking patterns, using visual appearance and motion cues across cameras and other sensors.

See all 29 papers

DiffCrossGait: Trajectory-Level Alignment for 2D-3D Cross-Modal Gait Recognition via Latent Diffusion
May 29, 2026Zhiyang Lu, Ming ChengDiffusion Model Alignment Latent Diffusion Models
1Fujian Key Laboratory of Urban Intelligent Sensing and Com · 2Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Min
Diffusion-Based Generation of Gait Trajectories
Sep 13, 2026Damian Benasco, Juan Carballeira-Lopez, Jaime Ramos-Rojas +4Trajectory Generation Diffusion Models for Motion Generation
Cajal Neuroscience Center, Spanish National Research Council (CSIC), Madrid, Spain · Bioengineering Systems and Technologies Research Group, Rey Juan Carlos University (URJC), Madrid, Spain
Vocabulary-Guided Gait Recognition
Sep 16, 2026Panjian Huang, Saihui Hou, Chunshui Cao +2Gait Recognition
School of Artificial Intelligence, Beijing Normal University

May 29, 2026J/K move · Enter open · S save

DiffCrossGait: Trajectory-Level Alignment for 2D-3D Cross-Modal Gait Recognition via Latent Diffusion

Zhiyang Lu, Ming Cheng

1Fujian Key Laboratory of Urban Intelligent Sensing and Com · 2Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Min

Diffusion Model Alignment Latent Diffusion Models Cross-Modal Alignment Gait Analysis Gait Recognition

Open paper AskPDF arXiv

Cross-modal 2D-3D gait recognition is impeded by inherent domain discrepancies between 2D silhouette and 3D LiDAR range-view representations. While prior methods align only final embeddings, we propose DiffCrossGait, which reformulates cross-modal matching as trajectory-level alignment in an identity-relevant latent diffusion space, rather than assuming full equivalence between 2D and 3D observations. By driving both modalities with shared Gaussian noise within a latent space, we enable continuous alignment throughout the generative evolution. We introduce a Tri-Phase Alignment Strategy that exploits varying noise intensities to enforce identity anchoring, dynamics consistency, and cross-modal structural recoverability, thereby constraining both modalities to share denoising dynamics and bottleneck structure, which promotes modality-invariant gait features. Crucially, our framework decouples generative alignment from the discriminative backbone; the diffusion mechanism serves exclusively as a training objective, ensuring high inference efficiency by eliminating the computational overhead of iterative denoising. Extensive experiments on the SUSTech1K and FreeGait benchmarks demonstrate that DiffCrossGait achieves state-of-the-art performance.

Abstract

Gait Recognition

Explore similar work

DiffCrossGait: Trajectory-Level Alignment for 2D-3D Cross-Modal Gait Recognition via Latent Diffusion

Diffusion-Based Generation of Gait Trajectories

Vocabulary-Guided Gait Recognition