Learning-AI
Vehicle Re-ID for Surround-view Camera System
June 2020
tl;dr: Vehicle reID with fisheye surround cameras.
Overall impression
Another practical work from Zongmu. The tracking
Key ideas
- single camera reID
- This should be largely tracking. –> why this is called ReID?
- SiamRPN++ model for single camera tracking
- multi camera reID
- BDBnet model (batch dropblock, ICCV 2019, SOTA for pedestrian ReID)
- association rule based on physics and geometry to only perform vehicle ReID in overlapping FoV
- feature distance and geometry constraint distance
- BDBnet model (batch dropblock, ICCV 2019, SOTA for pedestrian ReID)
Technical details
Notes
- Talk at CVPR 2020
- I feel that only the physical constraint will already do very good cross camera reID of the same vehicle. Do we really need reID module?