/WePerson

[ACM MM-2021] WePerson: learning a generalized re-identification model from all-weather virtual data

Primary LanguagePythonMIT LicenseMIT

WePerson

This repository contains the WePerson dataset proposed in our paper "WePerson: Learning a Generalized Re-identification Model from All-weather Virtual Data".

Fig. 1. Sample images from the proposed WePerson dataset.

Table of Contents

Dataset Description

The WePerson dataset is generated by GTA V with Script Hook V and scripthookvdotnet. This is the largest synthetic person re-identification dataset that includes 4,000,000 images from 1,500 identities. Images in the dataset are collected with 40 viewpoints per scene, various poses, complex backgrounds, occlusions, 7 different types of weather, 7 different types of natural illuminations. Identities in the dataset not only have different clothes but also wearable accessories, such as sunglasses, hats, etc. We show some dataset properties in Fig. 2.

Fig. 2. Properties of the proposed WePerson dataset.

Download Link

Due to the large amount of data, currently only a subset of WePerson is provided in Baidu Pan.

The subset can be downloaded from the link: BaiDu Pan

File Structure

WePerson Dataset
├── WePerson/
|   ├── 0000/
|   |   ├── 0000_c00_s00_T04.png
|   |   ├── 0000_c00_s00_T12.png
|   |   ├── 0000_c01_s03_T20.png
|   |   └── ...
|   ├── 0001/
|   |   ├── 0001_c03_s10_T10.png
|   |   ├── 0001_c06_s07_T20.png
|   |   ├── 0001_c10_s03_T4.png
|   |   └── ...
|   └── ...

The filenames are encoded as follows. Take "0000_c00_s00_T04.png" as an example,

  • 0000 is the id of the person
  • c00 is the id of the camera
  • s00 is the id of the scene
  • T04 is the current time that is related to natural illumination.

Experiment Results

By training person re-identification models on our dataset, we demonstrate that the model trained on virtual data outperforms the model trained on real-world dataset. Our dataset also surpasses other synthetic dataset in direct transfer.

In direct transfer evaluation, we achieve:

Citation

@inproceedings{li2021weperson,
  title={WePerson: Learning a Generalized Re-identification Model from All-weather Virtual Data},
  author={Li, He and Ye, Mang and Du, Bo},
  booktitle={29th ACM International Conference on Multimedia (ACMMM)},
  year={2021}
}

Contacts

He Li

School of Computer Science, Wuhan University

lihe404@whu.edu.cn