Overview

The Video-based Vehicle Re-Identification (VVeRI-901) is the first video-based vehicle re-identification(re-ID) dataset. The official full version of VVeRI-901 provides 901 vehicle IDs while the trial version includes 95 IDs.

Some distinctive characteristics of VVeRI-901 can be summarized as follows:

a) Unconstrained capture conditions involving multiple intersections motivate visual information diversity in view-point, resolution, and illumination, etc.
b) Successive spatial and temporal information with-out any further down-sampling is contained to enhance the appearance-based model's robustness in tackling visual ambiguities.
c) Besides the video-based vehicle re-ID task, with the aid of rich information, more related research areas can be facilitated, e.g., cross-resolution re-ID, cross-view matching and multi-view synthesis.

Data Statistics

The proposed dataset contains 901 IDs (451 IDs for training and 450 IDs for testing.), 2,320 tracklets, and 488,195 bounding boxes. All the IDs are captured by at least two cameras, and most of the IDs are captured by 2-4 cameras, indicating that VVeRI-901 is an ideal benchmark for algorithms to explore multiple queries or re-ranking methods. Generally, sequences from all cameras contain about 100-400 frames, guaranteeing information diversity in the proposed dataset.

Privacy Concern

We mask the license plates in the VVeRI-901 for privacy consideration and make the model more focus on the vehicle's visual appearance.

Data Collection

Sensor Deployment

The raw video data of the VVeRI-901 are captured from a mid-sized city of China with an area of 1.1km*2.4km. In this region, 43 surveillance cameras are deployed at different traffic intersections.

Raw Videos Collection

The raw videos are collected between 6:00 am to 6:00 pm.
1280*960 resolution
25 fps framerate

Data Preprocess

According to the non-overlapping scenario required by re-ID tasks, we choose the raw videos captured from 11 surveillance cameras (one at each intersection).

Data Annotation

Bounding Boxes

To acquire the tracklets of each vehicle in every single camera, we manually annotate the bounding box of each tracklet with the help of the Computer Vision Annotation Tool (CVAT).

ID Association

After getting all the tracklets in each camera, we associate the same vehicle that appears at different cameras with the auxiliary spatial-temporal cues.

ROI

It is worth noting that we find some vehicles stopping and keeping still at zebra crossings in some raw videos. These vehicles provide the same appearance in each frame and poss limited variations in a tracklet, leading to severe information redundancy in the dataset. In order to avoid this redundancy issue, we define a region of interest (ROI) for each scenario and retain the clips within the ROI.

Data Format

Folder Structure

VVeRI-901
├── Vehicle ID:1
│   ├── Camera ID:1
│   │   ├── 00001.png
│   │   ├── 00002.png
│   │   └── ...
│   ├── Camera ID:2
│   └── ...
├── Vehicle ID:2
├── Vehicle ID:3
...
├── Vehicle ID:901
├── train.txt
├── query.txt
└── gallery.txt

Label Format

The labels including vehicle IDs and camera IDs can be obtained from the folder names of the first-level directory and second-level directory of VVeRI-901.

- Vehicle ID: int
- Camera ID: string

Notes:

The first-level directory of the folder VVeRI-901 are all named by the vehicle IDs (e.g. '137').
The second-level directory of the folder VVeRI-901 are named by the camera IDs (e.g. 'L01_C01_V09'). Each camera ID offers the information of location (e.g., 'L01' in 'L01_C01_V09') and camera number (e.g., 'C01' in 'L01_C01_V09').
The files 'train.txt', 'query.txt' and 'gallery.txt' provide the directory of the tracklets (e.g., './VVeRI901_V1_Trial/137/L01_C01_V09').
Since we only provide the official trial version of the VVeRI-901, the IDs included are 95 instead of 901 and we split 45 IDs for trainning, 40 IDs for testing.

Citation

@Misc{VVeRI-901,
  title={VVeRI-901: Video Vehicle Re-Identification Dataset},
  author={Jianan Zhao, Fengliang Qi, Guangyu Ren and Lin Xu},
  note = {\url{https://www.graviti.cn/open-datasets/VVeRI901}},
  year={2020}
}