BarbieGait
收藏资源简介:
BarbieGait是由北京师范大学与阿里巴巴联合开发的合成步态识别数据集,通过虚拟引擎将真实受试者的步态特征映射至3D模型,实现身份一致性的跨服装模拟。该数据集包含521名受试者,每人100种随机服装组合(涵盖发型、上下装及配饰),生成120万条多模态序列(RGB图像、2D姿态、剪影),数据通过骨骼长度匹配、动态运动对齐等技术确保步态身份一致性。其高可控性的合成范式为跨服装步态识别研究提供了基准,旨在解决真实场景中服装多样性导致的类内方差增大问题,推动生物特征识别技术在安防监控等领域的应用。
BarbieGait is a synthetic gait recognition dataset jointly developed by Beijing Normal University and Alibaba. It maps the gait features of real subjects onto 3D models via virtual engines, enabling cross-clothing simulation with consistent identity. This dataset contains 521 subjects, each paired with 100 random clothing combinations covering hairstyles, upper/lower garments and accessories, generating 1.2 million multimodal sequences including RGB images, 2D poses and silhouettes. The data ensures consistent gait identity through techniques such as bone length matching and dynamic motion alignment. Its highly controllable synthetic paradigm provides a benchmark for cross-clothing gait recognition research, aiming to address the issue of increased intra-class variance caused by clothing diversity in real-world scenarios, and promote the application of biometric recognition technologies in fields such as security monitoring.
BarbieGait 数据集概述
基本信息
- 数据集名称:BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition
- 相关会议:CVPR 2026
- 作者:Qingyuan Cai, Saihui Hou, Xuecai Hu, Yongzhen Huang*
- 机构:北京师范大学人工智能学院 · 阿里巴巴集团AMAP · WATRIX.AI
- 项目主页:https://barbiegait.github.io/
- 论文预印本:https://arxiv.org/abs/2604.12221
数据集简介
BarbieGait是一个身份一致、支持多样化换装的合成人体步态识别数据集。
数据获取与使用
- 获取方式:用户需下载并签署《BarbieGaitGait Dataset Usage Agreement.pdf》协议,提交至BNU-IVC@outlook.com邮箱申请使用。
- 处理周期:申请通常在一周内处理。
- 数据存放路径:
your_path/BarbieGait_data/
数据下载与解压
-
从Google Drive下载数据至
BarbieGait_data/目录。 -
执行解压命令: bash cd your_path/BarbieGait_data/ tar -xvjf BarbieGait_predsil_pkl.tar.bz2
-
创建符号链接至代码目录: bash cd your_path/BarbieGait_CVPR26_release/BarbieGait ln -s your_path/BarbieGait_CVPR26_release/BarbieGait_data ./BarbieGait_data
数据预处理
目录结构
BarbieGait_data/ ├── BarbieGait_predsil_pkl/ # 原始数据 (personID/clothID-seqID) ├── thick_label_by_nakeddiffnorm_eqchg/ # 衣物厚度标签 └── P2_BarbieGait_predsil_pkl/ # 输出:重组后的数据 └── {subject_id}/ └── {cloth_type}/ └── {view_id}/ `-- {view_id}.pkl
预处理步骤
执行以下命令进行数据重组: bash cd BarbieGait/datasets python create_symlnk.py
输出格式
原始文件夹名称(如cloth00-00)被重组为thick{thick_value}-{seq_in_thick}-cloth00-00格式,其中:
thick_value:衣物厚度类别(0-9)seq_in_thick:该厚度类别内的序列索引
此重组按衣物厚度对序列进行分组,便于后续跨衣物研究。
待完成事项
- 发布论文链接
- 发布BarbieGait数据集
- 发布GaitCLIF代码库
- 发布预训练模型及配置
- 完善文档和使用示例

- 1BarbieGait: An Identity-Consistent Synthetic Human Dataset with Versatile Cloth-Changing for Gait Recognition北京师范大学·人工智能学院; 阿里巴巴集团·AMAP; WATRIX.AI · 2026年



