Awesome
This is an official repository for "Occluded Gait Recognition with Mixture of Experts: An Action Detection Perspective" (ECCV2024), providing a pioneering Occluded Gait database (OccGait, Prepared already) and the code of GaitMoE (Remained to be completed).
Highlights in OccGait
(1) It provides 101 identities with 8 camera views and over 80K sequences.
(2) Each identity walks under 4 different types of occluded scenarios, providing access to analyze the impact of occlusion in practical applications.
(3) Due to the privacy issue, OccGait currently only provides the silhouette version. Other data types, e.g., pose, remain to be considered. It is essential to note that it is ACADEMIC USE ONLY.
Introduction of OccGait
The OccGait is collected in an indoor gait recognition laboratory. There are 3 cameras (Cam1 of 0◦, Cam2 of 45◦, Cam3 of 315◦) with 1920 × 1080 resolution in a square area (8m x 8m). During the data collection process, the subjects follow a walking routine 1-2-3-4 as shown below. Overlapping camera views caused by the combination of 3 cameras and 4 walking directions are filtered out, and therefore gait sequences with 8 camera views are obtained. Four types of occluded scenarios are considered: None of Occlusion (Normal Walking as NM), Carrying Occlusion (CA), Crowd Occlusion (CR), and Static Occlusion (ST).
1. The Diagram of Data Collection Process
<img src="./assets/collection_process.png" width = "585" height = "270"/>2. The Diagram of 4 Types of Occlusion Scenarios
<img src="./assets/occlusion_scenarios.png" width = "830" height = "360"/>3. Examples of OccGait (RGB, Instance Segmentation with Mask2Former, and Silhouette)
<img src="./assets/rgb_seg_sil_masked.gif" width = "810" height = "270"/>- The Directory Structure of OccGait Dataset
001 (subject) CA01 (type) 000 (view) 0072.png (frame) 0073.png (frame) ...... ...... ......
Download OccGait
To obtain the OccGait dataset, all users are required to complete the following steps:
- Download and sign the latest agreement within this repository.
- Use the organization's email and send a request with a signed agreement attached to BNU-IVC_OccludedGait@outlook.com.
We will handle your requests within a week (Occasionally, emails may be flagged as spam. If you haven't received a response within a week, please resend your mail from an alternate email address.). In case you encounter any issues, please feel free to reach out to us via BNU-IVC_OccludedGait@outlook.com. <br>
GaitMoE
Remained to be completed.
Citation
Please cite the following paper if you find this useful in your research:
@InProceedings{Huang_2024_ECCV,
author = "Panjian, Huang and Yunjie, Peng and Saihui, Hou and Chunshui, Cao and Xu, Liu and Zhiqiang, He and Yongzhen, Huang",
title = "Occluded Gait Recognition with Mixture of Experts: An Action Detection Perspective",
booktitle = "Computer Vision -- ECCV 2024",
year = "2024"
}
Note: This dataset is only used for ACADEMIC PURPOSES, anyone can not use this dataset for anything that might be considered commercial use.
For OccCASIA-B dataset, please refer to OccSilGait.