CASIA-B Pose

Gait recognition as a prominent field within biometrics and computer vision focuses on analyzing the unique walking patterns of individuals for identification and security purposes. CASIA-B dataset encompasses a diverse collection of gait patterns with variations in walking styles, clothing, and viewpoints. This rich variety enables researchers to train and evaluate gait recognition algorithms on a comprehensive dataset that reflects real-world scenarios. Additionally, pose holds its own advantage in terms of robustness against carrying objects and clothing, making it particularly attractive for practical applications. However, most pose-based gait recognition methods tend to employ pose benchmarks calculated with different pose estimation algorithms, which led to duplicated work and biased comparisons. Hence, we introduce CASIA-B Pose, a pose-based gait recognition benchmark that utilizes 17 keypoints extracted by two state-of-the-art pose estimation algorithms, i.e. HRNet[1] and SimCC[2]. We also provide the mapping relations between RGB frames and keypoint tensors to foster further development. More details can be seen in [https://github.com/BNU-IVC/FastPoseGait].[1] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. "Deep High-Resolution Representation Learning for Visual Recognition," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 10, pp. 3349-3364, 1 Oct. 2021, doi: 10.1109/TPAMI.2020.2983686.[2] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. "SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation." Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI. Cham: Springer Nature Switzerland, 2022.If you use our dataset in your research, please cite the corresponding paper:Yang Fu1, Shibei Meng1, Saihui Hou* and Xuecai Hu and Yongzhen Huang*. "GPGait: Generalized Pose-based Gait Recognition." arXiv preprint arXiv:2303.05234 (2023). (The first two authors contribute equally to this work)@article{fu2023gpgait,title={GPGait: Generalized Pose-based Gait Recognition},author={Fu, Yang and Meng, Shibei and Hou, Saihui and Hu, Xuecai and Huang, Yongzhen},journal={arXiv preprint arXiv:2303.05234},year={2023}}

步态识别（Gait Recognition）作为生物特征识别（biometrics）与计算机视觉（Computer Vision）领域的重要分支，专注于分析个体独特的行走模式，以实现身份识别与安全保障目标。CASIA-B数据集涵盖了多样化的步态样本集，包含行走姿态、着装及拍摄视角等多种变化因素。这种丰富的样本多样性，使得研究者能够在贴合真实应用场景的全面数据集上训练并评估步态识别算法。
此外，基于人体姿态的步态识别方法在应对携带物品与着装变化时具备更强的鲁棒性，因此在实际应用中极具吸引力。然而，当前多数基于姿态的步态识别方法往往采用不同姿态估计（pose estimation）算法生成的姿态基准数据，这导致了重复研究与对比偏差的问题。
为此，我们提出CASIA-B Pose——一款基于姿态的步态识别基准数据集（benchmark dataset），其使用了由两种顶尖姿态估计算法（即HRNet[1]与SimCC[2]）提取的17个人体关键点（keypoints）。我们还提供了RGB帧（RGB frames）与关键点张量（keypoint tensors）之间的映射关系，以推动该领域的进一步发展。更多细节可参阅[https://github.com/BNU-IVC/FastPoseGait]。

[1] 王京东、孙柯、程天恒、姜博睿、邓朝瑞、赵阳、刘东、穆亚东、谭明奎、王兴刚、刘文予、肖斌："Deep High-Resolution Representation Learning for Visual Recognition"，发表于《IEEE模式分析与机器智能汇刊》，第43卷第10期，第3349-3364页，2021年10月1日，DOI: 10.1109/TPAMI.2020.2983686。
[2] 李延杰、杨森、刘培东、张守奎、王云晓、王志成、杨万口、夏树涛："SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation"，发表于《计算机视觉——ECCV 2022：第17届欧洲计算机视觉大会论文集（第六部分）》，以色列特拉维夫，2022年10月23日至27日，施普林格自然瑞士出版社，2022年。

若您在研究中使用本数据集，请引用以下论文：傅阳1、孟诗贝1、侯赛辉*、胡学才、黄勇珍*："GPGait: Generalized Pose-based Gait Recognition"，arXiv预印本arXiv:2303.05234（2023）。（本文前两位作者贡献均等）
对应的BibTeX引用格式如下：
@article{fu2023gpgait,
title={GPGait: Generalized Pose-based Gait Recognition},
author={Fu, Yang and Meng, Shibei and Hou, Saihui and Hu, Xuecai and Huang, Yongzhen},
journal={arXiv preprint arXiv:2303.05234},
year={2023}
}

搜集汇总

背景与挑战

背景概述

CASIA-B Pose是一个基于姿态的步态识别基准数据集，它利用HRNet和SimCC两种先进算法提取了17个关键点，以解决现有方法中因使用不同姿态估计算法而导致的重复工作和有偏比较问题。该数据集提供了RGB帧与关键点张量的映射关系，支持步态识别算法的训练和评估，特别强调姿态在应对携带物品和衣物变化方面的鲁棒性，适用于实际安全应用场景。