we show simple reproduction of circle loss for reid. we let resnet50 as the baseline without any trick.
paper: https://arxiv.org/pdf/2002.10857.pdf
>>> python train2.py
- Dataset: market1501
- baseline: resnet50 without fc and tricks
| rank-1 | mAP | gamma | m | lr | lr-decay | step | batch |
|---|---|---|---|---|---|---|---|
| 82% | 56 | 0.3 | 0.00035 | 0.1 | 30 | 128 |
the dist of the circle loss achieved by cos dist in paper, we add euclidean dist in the loss.py
The codes are expanded on
- ReID-baseline by Luo & Liao