OU-ISIR Gait Database, Multi-View Large Population Dataset with Human Meshes

Introduction

The OU-ISIR Gait Database, Multi-View Large Population Dataset with Human Meshes (OUMVLP-Mesh) is meant to aid research efforts in the general area of developing, testing and evaluating algorithms for model-based gait analysis. The Institute of Scientific and Industrial Research (ISIR), Osaka University (OU) has copyright in the collection of gait video and associated data and serves as a distributor of the OUMVLP-Mesh dataset.

Existing model-based gait databases provide the 2D poses (i.e., joint locations) extracted by general pose estimators as the human model. However, these 2D poses suffer from information loss and are of relatively low quality. Therefore, OUMVLP-Mesh considers a more informative 3D human mesh model with parametric pose and shape features (i.e., SMPL model). The SMPL model is complex and informative, making it conducive to promoting the development of model-based gait recognition and boosting the recognition performance. Besides gait recognition, OUMVLP-Mesh can also be used for other gait analyses (e.g., aging progression/regression, training data as genuine gait models for adversarial learning). If you use this dataset please cite the following paper:

  • X. Li, Y. Makihara, C. Xu, Y. Yagi, "Multi-view Large Population Gait Database with Human Meshes and Its Performance Evaluation", IEEE Transactions on Biometrics, Behavior, and Identity Science (T-BIOM), vol. 4, no. 2, pp. 234-248, April 2022. [Bib]

Dataset

Built upon OU-MVLP, OUMVLP-Mesh contains 10,307 subjects with up to 14 viewing angles (from 0° to 90°, 180° to 270° at 15° intervals). The raw images are captured at a resolution of 1280 x 980 pixels at 25 fps under the camera setup shown in Fig. 1. There are seven network cameras placed at 15° intervals along a quarter of a circle. The subjects walk from location A to location B and back to location A, producing 14 view sequences.



Fig.1: Camera setup and the estimated human mesh models from multiple views.

The human mesh models are then estimated from OU-MVLP using the proposed mesh estimation method. The estimated models from all 14 views are shown in Fig. 1. Some detailed examples are shown in Fig. 2. Besides the 3D meshes, the corresponding skeletons are also provided, and have two different sets of coordinates. The image-based 2D (IM2D) coordinates are the same as those of the poses in OUMVLP-Pose; however, 1D information has been lost and the coordinates change with the viewing angle of the input image. The human-centered 3D (HC3D) coordinates recover the missing information of the input 2D images and are aligned with a common coordinate that is robust to the viewing angle. There are 24 joints in the skeleton shown in Fig. 3, including Pelvis, Lhip, Rhip, Spine1, Lknee, Rknee, Spine2, Lankle, Rankle, Spine3, Lfoot, Rfoot, Neck, Lcollar, Rcollar, Head, Lshoulder, Rshoulder, Lelbow, Relbow, Lwrist, Rwrist, Lhand, and Rhand.



Fig.2: Examples from two different views. The first row is the cropped RGB sequences; the second, third, and fourth rows are the corresponding human mesh models, 2D image-based joint locations, and 3D human-centered joint locations, respectively.



Fig.3: Human skeleton model with 24 joints.

The number of frames in a sequence is from 18 to 35, and most of the sequences contain approximately 25 frames.

How to get the dataset?

To advance the state-of-the-art in gait-based application, this dataset including a set of sequences (i.e., IM2D joints, HC3D joints, 85-D SMPL parameters, and HC vertices) and subject ID lists of training and testing set could be downloaded as a zip file with password protection and the password will be issued on a case-by-case basis. To receive the password, the requestor must send the release agreement signed by a legal representative of your institution (e.g., your supervisor if you are a student) to the database administrator by mail, e-mail, or FAX.

  • Release agreement
  • Dataset (All data is separated into several zip files. In each zip file, the naming rule is ./{subject}/{view}_{sequence}/{frame}.json):
  • Synthesized multi-view gait dataset containing the cropped RGB sequences and silhouettes and GT 3D pose

    • The database administrator

    Department of Intelligent Media, The Institute of Scientific and Industrial Research, Osaka University
    Address: 8-1 Mihogaoka, Ibaraki, Osaka, 567-0047, JAPAN
    Mail address
    FAX: +81-6-6877-4375.