Repository for 2019 CVPR AI City Challenge Track 2 from IPL @University of Washington. Our method ranks 2nd in the competition.
Our code consists of the following three components:
The multi-view and metadata re-ranking vehicle reidentification model. The code is based on Jiyang Gao's Video-Person-ReID [code].
Metadata model for vehicle's type, brand and color. The code is based on [code].
The vehicle keypoints code is based on krrish94's CarKeypoints [code].
Training of both Video-Person-ReID and metadata requires CarKeypoints's inference result on training set. For CarKeypoints, we use the pre-trained model [model]. Please refer to the README.md files in each subfolder.
Testing of both Video-Person-ReID and metadata requires CarKeypoints's inference result on testing set. In addition, Video-Person-ReID needs metadata's inference result on testing set.