KY 4D Gait Database A (Straight)

Introduction

Kyushu University 4D Gait Database A (Straight) is composed of sequential 3D models and image sequences of people walking along straight trajectories. This database comprises 42 subjects with four sequences for each subject. All people walked along straight trajectories, as indicated by dashed line 1 in the figure. Multiple 3D models were reconstructed by the visual hull technique with 16 cameras placed in the studio. This dataset was first introduced in the EST 2010 paper, Person identification from spatio-temporal 3D gait" [1]. Detailed analysis is available in the paper in Pattern Recognition Letters 2014 Identification of people walking along curved trajectories" [2].

Example sequential 3D models of a person walking straight

Dataset

The KY 4D Gait Database A consists of sequential 3D models of forty two walking people, color and silhouette images taken by 16 cameras, and camera parameters of each camera. For more details, please refer to [1].

Download

You can download dataset from anonymous FTP server (ftp://robotics-ftp.ait.kyushu-u.ac.jp). All images and 3D models are in "gait" folder.

Citation

If you make use of the KY 4D Gait Database A in any form, please do cite the following paper [1]:

[1] Y. Iwashita, R. Baba, K. Ogawara, and R. Kurazume, "Person identification from spatio-temporal 3D gait", EST 2010.

[2] Y. Iwashita, K. Ogawara, and R. Kurazume, "Identification of people walking along curved trajectories", Pattern Recognition Letters, to appear, 2014.

@inproceedings{yumi2010gait,
title={Person identification from spatio-temporal 3D gait},
author={Y. Iwashita and R. Baba and K. Ogawara and R. Kurazume},
booktitle={Int. Conf. Emerging Security Technologies (EST)},
page={30-35},
year={2010},
month={September},
address={UK},
}

List of performance

Here is a list of performance presented in papers. If you have new results and want to show them here, please email me (yumi@ieee.org) with your paper information.

1. Following methods assume that the gallery dataset contains 3D models (i.e. multiple viewpoint images), but the probe dataset contains single viewpoint images.

Year	Information about targets	Performance [%]	Paper information

2013	21 people, walking on straight path, 2D image-based method	- 90.5 (camera angle: 0 degree, frontal view images) - 85.7 (camera angle: 90 degree, side view images)	Expanding gait identification methods from straight to curved trajectories, Yumi Iwashita, Koichi Ogawara, Ryo Kurazume, IEEE Winter Conference on Applications of Computer Vision, pp.193-199, Jan 17-18. 2013.

2014	42 people, walking on straight path, 2D image-based method	- 99.4 (camera angle: 0 degree, frontal view images) - 96.4 (camera angle: 45 degree) - 98.2 (camera angle: 90 degree, side view images)	Identification of people walking along curved trajectories, Yumi Iwashita, Koichi Ogawara, Ryo Kurazume, Pattern Recognition Letters, Vol. 46, pp. 60-69, 2014. (DOI: 10.1016/j.patrec.2014.04.004)

2. Following methods assume that both gallery and probe datasets contain 3D models (i.e. multiple viewpoint images).

Year	Information about targets and methodology	Performance [%]	Paper information

2015	42 people, walking on straight path, 3D model-based method	- 95.13 (averaged rank-1 performance)	Entropy volumes for viewpoint-independent gait recognition, D. Lopez-FernandezEmail authorF. J. Madrid-CuevasA. Carmona-PoyatoR. Munoz-SalinasR. Medina-Carnicer, Machine Vision and Applications, Volume 26, Issue 7, pp 1079-1094, 2015.

2016	42 people, walking on curved path, 3D model-based method	- 99.4 (averaged rank-1 performance)	A new approach for multi-view gait recognition on unconstrained paths, D. Lopez-Fernandez, F.J. Madrid-Cuevas, A. Carmona-Poyato, R. Munoz-Salinas, R. Medina-Carnicer, Journal of Visual Communication and Image Representation Volume 38, July 2016, Pages 396-406, 2016.

Updated 11/14/2016