pkuvcl/pku-vehicles

PKU-Vehicle

Group Sensitive Triplet Embedding for Vehicle Re-identification

Description of VERI-Wild Dataset

To meet the emerging demand on large-scale vehicle re-identification, we construct a dataset, namely PKU-Vehicle, which contains tens of millions of vehicle images captured by real-world surveillance cameras in several cities in China. The PKU-Vehicle dataset contains 10 million vehicle images captured from multiple real video surveillance systems across several cities, which serves as a distractor dataset to test the large-scale retrieval performance. Various locations (\textit{e.g.}, highways, streets, intersections), weather conditions (\textit{e.g.}, sunny, rainy, foggy), illuminations (\textit{e.g.}, daytime and evening), shooting angles (\textit{e.g.}, front, side, rear), different resolutions (\textit{e.g.}, 480P, 640P, 720P, 1080P, 2K) and hundreds of vehicle brands are involved in PKU-Vehicle dataset.

PKU-Vehicle dataset is collected from different surveillance cameras with 10 millions images. The vehicle objects in images are cropped out, such that each image contains one vehicle. In order to thoroughly evaluate the re-identification methods at different scales, we further split the database into eight subsets, i.e., 10 thousands, 50 thousands, 100 thousands, 500 thousands, 1 million, 2 millions, 5 millions, 10 millions.

Download Links

Baidu Yun Pan: https://pan.baidu.com/s/198hdpC0NvXfboxOYjT_haw Password: cpql

Google Driver: https://drive.google.com/drive/folders/12gV4gOA4C8FbFzL5t3oJbDxS8_LMDyis?usp=sharing

Citation

@article{bai2018group,
  title={Group-sensitive triplet embedding for vehicle reidentification},
  author={Bai, Yan and Lou, Yihang and Gao, Feng and Wang, Shiqi and Wu, Yuwei and Duan, Ling-Yu},
  journal={IEEE Transactions on Multimedia},
  volume={20},
  number={9},
  pages={2385--2399},
  year={2018},
  publisher={IEEE}
}