The PersonX dataset is generated by the synthetic data engine PersonX that can synthesise images under controllable cameras and environment.
For researching viewpoint, six backgrounds are used in "Dissecting Person Re-identification from the Viewpoint of Viewpoint" (as shown in the following figure), including three pure color backgrounds and three scene backgrounds. There are 1266 hand-crafted identities (547 females and 719 males) and each identety has 36 viewpoints (sampled every from to ).
To provide data for “synthetic to real” study, we generate more diverse data and build two new datasets. One is used in ViSDA2020 Chanllenge and the other one is used in Alice Benchmakrs. The main difference between the two data sets is that the environmental factor parameters of the latter are more complex.
3D identetie models of PersonX and images of above 6 backgrounds can be downloaded from the following links:
Directories & Files of images
PersonX_v1
├── 1/ # this number is coresponds with the background shown in above figure.
│ ├── bounding_box_test/
│ │ ├── 0001_c1s1_01.jpg
│ │ ├── 0001_c1s1_02.jpg
│ │ └── 0001_c1s1_03.jpg
│ ├── bounding_box_train/
│ │ ├── 0004_c1s1_01.jpg
│ │ ├── 0004_c1s1_02.jpg
│ │ └── 0004_c1s1_03.jpg
│ └── query/
│ ├── 0001_c1s1_32.jpg
│ ├── 0002_c1s1_02.jpg
│ └── 0003_c1s1_15.jpg
├── 2~6/
│ └── ...
└── readme.txt
Name of the image
Taking "0001_c1s1_33.jpg" as an example:
Directories & Files of images are similar with PersonX_v1.
Coming Soon!
The package can be imported into any program of Unity directly.
We have supplied a example of a controller "example-controllerc.unitypackage" that can make the person to rotate and can capture the images automatically.
Components in Game_Manager_Demo
Game_Manager_Demo
├── PeoplePosition # control the position, rotation interval of the person
│
│
├── Camara # capture images of person without background
│
│
└── Main Camera # capture images of person
Main Camera and Camera are used to generate a pair of images that have cintains the same person in same position. One is without background, which is used to caculate the bounding box of the person:
!! the property of people and two cameras:
The layer of all person is role;
The culling mask of Camera is role (only),so images captured by Camera contains the person without background;
The culling mask of Main Camera is Everything.
!! two paths:
The name of the path containing the person models shoud be "Resources" (iyou can directly import our unitypackage);
The 24th line of Manager.cs is the path to save the generated images, which is better not to be the sub-path of the project (because Unity will import the generating images, which will slow the generation process.).
!! the ID of the first person:
!! the version of Unity:
If you use this dataset in your research, please kindly cite our work as,
@inproceedings{sun2019dissecting,
title={Dissecting Person Re-identification from the Viewpoint of Viewpoint},
author={Sun, Xiaoxiao and Zheng, Liang},
booktitle={CVPR},
year={2019}
}