Large synthetic group re-identification dataset, including over 1M images.

The proposed City1M is constructed by collecting images from 8 street cameras in a city scene, which contains 1.84M RGB images with a uniform resolution of 1920 x1080, 45,000 persons, and 11,000 groups. Our dataset also includes day and night scenes.

We provide a full version of the City1M (named City1M-Full), a simplified version (named City1M-Lite), and related person models (named City1M-PersonModels).

The City1M-full consists of 6 zip files, named Images_x.zip, x=2,3,4,5,6. For example, the format of image name in each zip file is:

Images_2/0009501_pCount2_cam1_0000001_day.jpg

• '00095001' is the group id of this image.
• 'pCount2' is the number of member. (this is only used for creating datraset, maybe not useful for you.)
• 'cam1' is the camera id.
• '00000001' is the index. (it also makes no sense for you.)
• 'day' is a coarse time label (totally two options, 'day' or 'night').

We provide two protocols as follows:

• Procotol@1 = train_part1.txt + query_part1.txt + gallery_part1.txt
• Procotol@2 = train_part2.txt + query_part2.txt + gallery_part2.txt

Each line in txt file is a label of the image:

0009501_pCount2_cam1_0000001_day.jpg 0,1477.3156,230.6843,186.0616,346.3864,1.5166 1,1752.9762,215.4119,153.7662,221.532,2.6906

Each line is split by ' ', the first item is the image name, and the following items are the annotations for each member, for example:

0,1477.3156,230.6843,186.0616,346.3864,1.5166

• 0 is the person id of the member.
• 1477.3156,230.6843,186.0616,346.3864 is the (x,y,h,w) of the member.
• 1.5166 is depth between the camera and the member.

Considering that it may be difficult to download City1M-full directly, because the original dataset contains more than one million images with a uniform resolution of 1920 x1080 (~405G after unzip), we also provide a simple version, City1M-lite.

City1M-Lite has cropped the group members in the image, and provides 16 images of the corresponding background (8 cameras with 'day' and 'night').

For example, 'Images_2/0009501_pCount2_cam1_0000001_day.jpg' in City1M-full is formatted as 'Image_2_crop/0009501/{0.jpg, 1.jpg}' in City1M-lite.

If you need the whole image with a low download cost, you can reconstruct it using corresponding txt files and background images.

We also provided the pedestrian 3D models used in the City1M. We produced 46,000 models for our dataset, in which 45,000 models with IDs between 1 and 45,000 are actually used to construct the dataset, and the rest 1,000 ones are for standby. Each pedestrian model is an FBX file, together with texture files in the Texture directory.

If you are interested in our dataset, please feel free to fill in the application form and send it to email [email protected], and we will reply to you as soon as possible.

If you only need to obtain the 3D coordinates of members from the group images of other GReID datasets, we recommend that you use the following monocular depth estimation projects and algorithms. https://github.com/aim-uofa/AdelaiDepth/tree/main/LeReS

@InProceedings{Zhang_2022_CVPR,
    author    = {Zhang, Quan and Dang, Kaiheng and Lai, Jian-Huang and Feng, Zhanxiang and Xie, Xiaohua},
    title     = {Modeling 3D Layout for Group Re-Identification},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {7512-7520}
}