BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition

Qingyuan Cai; Saihui Hou; Xuecai Hu; Yongzhen Huang

cs.CVApr 14, 2026

BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition

Authors: Qingyuan Cai, Saihui Hou, Xuecai Hu, Yongzhen Huang

Organizations: School of Artificial Intelligence, Beijing Normal University · AMAP, Alibaba Group

Abstract

Gait recognition, as a reliable biometric technology, has seen rapid development in recent years while it faces significant challenges caused by diverse clothing styles in the real world. This paper introduces BarbieGait, a synthetic gait dataset where real-world subjects are uniquely mapped into a virtual engine to simulate extensive clothing changes while preserving their gait identity information. As a pioneering work, BarbieGait provides a controllable gait data generation method, enabling the production of large datasets to validate cross-clothing issues that are difficult to verify with real-world data. However, the diversity of clothing increases intra-class variance and makes one of the biggest challenges to learning cloth-invariant features under varying clothing conditions. Therefore, we propose GaitCLIF (Gait-oriented CLoth-Invariant Feature) as a robust baseline model for cross-clothing gait recognition. Through extensive experiments, we validate that our method significantly improves cross-clothing performance on BarbieGait and the existing popular gait benchmarks. We believe that BarbieGait, with its extensive cross-clothing gait data, will further advance the capabilities of gait recognition in cross-clothing scenarios and promote progress in related research.

Figures & tables

Figure 2: The BarbieGait data generation system includes: (a) Skeleton Length and Body Shape Matching maps real humans to virtual ones based on 3D skeleton and body shape using MakeHuman [ 47 ] . (b) Random Dressing refers to randomly selecting outfits for cloth-changing. (c) Kinematic Motion Matching refers to the alignment of gait identity information across different outfits of the same real subject. (d) Scene Construction uses Blender [ 12 ] to create multiple environments and capture multi-view images. (e) Rendering refers to accelerate image generation using a GPU cluster.

	Dataset	Year	Type	#Subj .	#Mesh	#Views	#Seq	Cloth-Changing	#Cloth/Subj.	GT Format
Motion	SURREAL [ 60 ]	2017	Synthetic	—	145	1	NA	$\times$	—	3DM
	AGORA [ 51 ]	2021	Synthetic	—	>350	1	NA	$\times$	—	3DM, silh.
	GTA-Human [ 8 ]	2021	Synthetic	—	>600	1	20K	$\times$	—	3DM
	SynBody [ 73 ]	2023	Synthetic	—	10,000	4	40K	$\times$	—	3DM, silh.
Re-ID	RandPerson [ 64 ]	2020	Synthetic	—	8,000	19	NA	$\times$	—	NA
	UnrealPerson [ 80 ]	2021	Synthetic	—	3,000	34	NA	$\times$	—	NA

Table 1: Comparison with existing Motion, Person Re-ID, and Gait datasets. “#Subj” denotes the number of subjects corresponding to real identities. “#Mesh”, “#Views”, and “#Seq” represent the total number of meshes, camera viewpoints, and sequences, respectively. “Cloth-Changing” indicates whether subjects appear in multiple outfits, and “#Cloth/Subj.” specifies the number of clothing variations per subject. “GT Format” lists available ground-truth types: “3DJ” (3D joints), “3DM” (3D meshes), and “silh.” (silhouettes).

Figure 3: Clothing Complexity and Thickness: (a) Silhouette without clothes. (b) Silhouette with clothes. (c) Non-overlapping area between (a) and (b) indicates garment complexity. (d) Distribution of subjects across thickness levels.

Figure 4: Overview of GaitCLIF. (a) GON, the core normalization unit. (b) GON-P3D and (c) GON-3D, two GON-based visual blocks used in the visual stages of GaitCLIF. (d) GON-FC, a GON-enhanced FC block used in the Head of GaitCLIF. (e) The overall GaitCLIF framework for cross-clothing gait recognition.

Dataset	Batch Size	Blocks	Milestones	Total
BarbieGait	(8,8)	[1,1,1,1]	(20k,40k,50k)	60k
CCPG [ 40 ]	(8,16)	[1,1,1,1]	(20k,40k,50k)	60k
SUSTech1K [ 55 ]	(8,8)	[1,1,1,1]	(20k,40k,40k)	50k
Gait3D [ 81 ]	(32,4)	[1,4,4,1]	(20k,40k,50k)	60k
GREW [ 82 ]	(32,4)	[1,4,4,1]	(80k,120k,150k)	180k

Table 2: Dataset-specific configurations including backbone and training settings.

Methods	THK1		THK2		THK3		THK4		THK5		THK6		THK7		THK8		THK9		AVG
Methods	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP
GaitSet [ 10 ]	20.6	23.3	15.8	18.9	12.7	15.9	12.3	15.5	7.4	10.9	6.0	8.9	5.3	7.8	4.1	7.5	3.2	6.2	9.7	12.8
GaitPart [ 21 ]	50.4	47.7	42.6	41.6	35.2	35.6	33.9	34.5	24.8	26.8	19.1	21.7	14.8	17.3	16.7	19.0	11.3	14.3	27.6	28.7
GaitGL [ 42 ]	44.7	34.0	37.5	29.4	31.1	25.2	31.2	25.7	22.3	19.5	18.3	16.6	13.0	13.2	14.8	14.6	13.8	13.8	25.2	21.3
GaitBase [ 19 ]	32.1	32.9	27.7	29.2	22.8	24.7	21.0	23.1	16.2	19.2	13.0	16.1	10.6	13.5	11.9	15.1	9.6	12.6	18.3	20.7
DeepGaitV2-2D [ 18 ]	59.6	57.9	51.6	51.3	46.1	45.5	42.6	43.7	36.4	38.2	29.3	31.9	22.5	25.5	25.5	28.7	18.0	21.9	36.8	38.3

Table 3: Performance comparison on BarbieGait when using predicted silhouette and 2D pose as input. Best result is in bold, and the second-best result is underlined. GaitCLIF-P3D ∗ uses heatmaps as input, following the same data processing pipeline as SkeletonGait. In our experiments, THK0 serves as the gallery and THK1-THK9 as probes.

GON-P3D	GON-FC	AVG-R1 (%)	AVG-mAP (%)
$\times$	$\times$	67.7	57.6
$\surd$	$\times$	69.8	57.6
$\times$	$\surd$	69.2	59.1
$\surd$	$\surd$	75.6	63.2

Table 4: Ablation studies for GaitCLIF-P3D on BarbieGait.

Methods	CCPG								SUSTech1K
	CL		UP		DN		MEAN		NM	BG	CL	CA	UM	UN	OC	NG	Overall
	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1								R1	R5
GaitSet [ 10 ]	77.7	46.4	83.5	59.6	83.2	61.4	81.5	55.8	69.1	68.2	37.4	65.0	63.1	61.0	67.2	23.0	65.0	84.8
GaitPart [ 21 ]	77.8	45.5	84.5	63.1	83.3	60.1	81.9	56.2	62.2	62.8	33.1	59.5	57.2	54.8	57.2	21.7	59.2	80.8
GaitBase [ 19 ]	88.5	—	92.7	—	93.4	—	91.5	—	81.5	77.5	49.6	75.8	75.5	76.7	81.4	25.9	76.1	89.4
SkeletonGait [ 20 ]	52.4	20.8	65.4	35.8	72.8	40.3	63.5	32.3	67.9	63.5	36.5	61.6	58.1	67.2	79.1	50.1	63.0	83.5

Table 5: Performance comparison on CCPG and SUSTech1K. For clarity, DeepGaitV2 and GaitCLIF refer to P3D-based models.

Methods	Gait3D			GREW
Methods	R1	R5	mAP	R1	R5
GaitSet [ 10 ]	36.7	58.3	30.0	46.3	63.6
GaitPart [ 21 ]	28.2	47.6	21.6	44.0	60.7
GaitBase [ 19 ]	60.1	—	—	64.6	—
GaitMoE [ 30 ]	73.7	—	66.2	79.6	89.1
DeepGaitV2 [ 18 ]	74.4	88.0	65.8	77.7	88.9

Table 6: Performance comparison on Gait3D and GREW. For clarity, DeepGaitV2 and GaitCLIF refer to P3D-based models.

Model	CCPG
	CL		UP		DN		MEAN
	R1 & mAP
DeepGaitV2	90.5	63.6	95.3	81.1	92.9	79.4	92.9	74.7
GaitCLIF (Scratch)	93.4	66.4	97.4	84.3	93.5	80.3	94.8	77.0
GaitCLIF (Pretrain)	93.9	69.8	97.9	85.6	95.7	83.0	95.8	79.5

Table 7: Cross-domain performance on CCPG using different training strategies. “Scratch” indicates training only on CCPG, while “Pretrain” denotes BarbieGait pretraining followed by CCPG fine-tuning. GaitCLIF refer to P3D-based models.

Methods	CCPG
	CL		UP		DN		MEAN
	R1 & mAP
SkeletonGait [ 20 ]	52.4	20.8	65.4	35.8	72.8	40.3	63.5	32.3
DPGait [ 37 ]	70.7	—	82.4	—	84.2	—	79.1	—
Ours	76.9	40.0	84.9	56.7	87.2	57.7	83.0	51.5

Table 8: BarbieGait improves upstream pose estimation, enabling better generalization to real-world dataset CCPG.

Figure 5: The pose format we used in our experiments. (a) COCO-17 format (b)Barbie-17 format.

Methods	Gait Recognition		Video Re-ID
Methods	HybridGait	OU-ISIR	CCVID	MEVID
DeepGaitV2 [ 18 ]	55.34	90.54	93.18	69.94
GaitCLIF (ours)	60.11	94.04	97.08	74.05

Table 9: Performance comparison of DeepGaitV2 and GaitCLIF on additional cloth-changing benchmarks.

GON-P3D	GON-FC	THK1		THK2		THK3		THK4		THK5		THK6		THK7		THK8		THK9		AVG
GON-P3D	GON-FC	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP
$\times$	$\times$	85.4	72.4	81.0	68.8	76.7	64.8	76.1	64.4	72.8	59.9	63.1	54.2	50.9	44.8	55.3	46.9	47.9	42.3	67.7	57.6
$\surd$	$\times$	86.1	71.3	81.8	68.0	78.3	64.3	78.2	64.2	75.3	59.8	64.7	54.2	53.6	45.8	58.9	47.2	51.2	43.6	69.8	57.6
$\times$	$\surd$	83.8	71.7	80.1	68.6	76.8	65.2	76.9	65.2	73.1	60.7	65.8	56.6	53.4	46.8	59.7	50.3	53.3	46.4	69.2	59.1
$\surd$	$\surd$	88.1	74.2	84.8	71.5	82.0	68.3	82.3	68.5	80.1	65.1	71.6	60.5	62.8	53.5	67.5	55.1	61.1	51.9	75.6	63.2

Table 10: Ablation study of each module under different clothing conditions (THK1-THK9). We report Rank-1 (R1) accuracy and mean Average Precision (mAP) for each variant.

Norm Type	THK1		THK2		THK3		THK4		THK5		THK6		THK7		THK8		THK9		AVG
Norm Type	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP
BN	83.2	71.8	78.6	68.0	74.2	63.8	73.6	63.2	69.1	58.3	60.9	53.6	47.8	43.0	54.7	46.9	45.9	41.6	65.3	56.7
IN	75.4	67.2	71.0	63.7	66.2	59.9	65.1	58.9	61.0	55.5	52.3	49.1	43.0	41.6	46.6	44.3	38.0	37.5	57.6	53.1
LN	76.9	61.7	71.9	58.4	66.7	54.6	66.3	54.3	61.9	50.4	53.5	45.8	43.3	38.4	50.1	42.0	41.1	35.8	59.1	49.0
GON	88.1	74.2	84.8	71.5	82.0	68.3	82.3	68.5	80.1	65.1	71.6	60.5	62.8	53.5	67.5	55.1	61.1	51.9	75.6	63.2

Table 11: Comparison of common normalization methods (BN, IN, LN) and our proposed GON across different clothing thickness levels.

Figure 6: The Illustration of our diverse clothing. BarbieGait includes a variety of hairstyles, clothing, shoes, and carried objects, introducing significant clothing variations for gait recognition under cloth-changing conditions.

Methods	Pose Type	Upstream Model	CCPG								SUSTech1K
			CL		UP		DN		Mean		NM	BG	CL	CA	UM	UN	OC	NG	Overall
			R1 & mAP								R1								R1	R5
GaitTR [ 77 ]	COCO-17	HRNet [ 63 ]	24.3	9.7	28.7	16.1	31.1	16.4	28.0	14.1	33.3	31.5	21.0	30.4	22.7	34.6	44.9	23.5	72.6	56.0
GaitGraph [ 59 ]	COCO-17	HRNet [ 63 ]	5.0	2.4	5.7	4.0	7.3	4.2	6.0	3.5	22.2	18.2	6.8	18.6	13.4	19.2	27.3	16.4	18.6	40.2
SkeletonGait [ 20 ]	COCO-17	HRNet [ 63 ]	52.4	20.8	65.4	35.8	72.8	40.3	63.5	32.3	55.0	51.0	24.7	49.9	42.3	52.0	62.8	43.9	50.1	72.6
DPGait [ 37 ]	COCO-17	ViTPose [ 71 ]	70.7	—	82.4	—	84.2	—	79.1	—	—	—	—	—	—	—	—	—	—	—

Table 12: BarbieGait improves upstream pose estimation, enabling better generalization to real-world dataset CCPG and SUSTech1K.

Figure 7: The illustration of our synthesized images. Our synthetic images are rendered in different scenes, realistic lighting conditions, diverse clothing conditions, and natural occlusions.

Figure 8: Visualization of heatmaps in Silhouette-based (a)-(c) and Pose-based methods (d)-(f). (b) and (e) show activation heatmaps of DeepGaitV2 and SkeletonGait overlaid on the silhouette. (c) and (f) show the effect with GaitCLIF.

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

MMGait: Towards Multi-Modal Gait Recognition
Apr 17, 2026Chenye Wang, Qingyuan Cai, Saihui Hou +2Cross-Modal Representation Learning Gait Analysis
School of Artificial Intelligence, Beijing Normal University
Who Remains, What Changes: Identity Anchored Composed Gait Retrieval
Aug 27, 2026Jingchen Fei, Zengbin Wang, Yukun Liu +3Cross-Modal Retrieval Gait Recognition
Beijing University of Posts and Telecommunications · Huazhong University of Science and Technology
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

Apr 17, 2026J/K move · Enter open · S save

MMGait: Towards Multi-Modal Gait Recognition

Chenye Wang, Qingyuan Cai, Saihui Hou, Aoqi Li+1

School of Artificial Intelligence, Beijing Normal University

Cross-Modal Representation Learning Gait Analysis Gait Recognition Multimodal Sensor Fusion

Open paper AskPDF arXiv

Gait recognition has emerged as a powerful biometric technique for identifying individuals at a distance without requiring user cooperation. Most existing methods focus primarily on RGB-derived modalities, which fall short in real-world scenarios requiring multi-modal collaboration and cross-modal retrieval. To overcome these challenges, we present MMGait, a comprehensive multi-modal gait benchmark integrating data from five heterogeneous sensors, including an RGB camera, a depth camera, an infrared camera, a LiDAR scanner, and a 4D Radar system. MMGait contains twelve modalities and 334,060 sequences from 725 subjects, enabling systematic exploration across geometric, photometric, and motion domains. Based on MMGait, we conduct extensive evaluations on single-modal, cross-modal, and multi-modal paradigms to analyze modality robustness and complementarity. Furthermore, we introduce a new task, Omni Multi-Modal Gait Recognition, which aims to unify the above three gait recognition paradigms within a single model. We also propose a simple yet powerful baseline, OmniGait, which learns a shared embedding space across diverse modalities and achieves promising recognition performance. The MMGait benchmark, codebase, and pretrained checkpoints are publicly available at https://github.com/BNU-IVC/MMGait.

Abstract

Gait Recognition

Explore similar work

MMGait: Towards Multi-Modal Gait Recognition

Who Remains, What Changes: Identity Anchored Composed Gait Retrieval

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