DR Loss
PyTorch Implementation for Our CVPR'20 Paper: "DR Loss: Improving Object Detection by Distributional Ranking"
Requirements
- Python 3.7
- PyTorch 1.1
- maskrcnn-benchmark
Usage:
- Put the loss file to the codebase of maskrcnn_benchmark at
maskrcnn-benchmark/maskrcnn_benchmark/layers/sigmoid_dr_loss.py
and add the class into "init.py".
- Change the focal loss in RetinaNet to the dr loss at
maskrcnn-benchmark/maskrcnn_benchmark/modeling/rpn/retinanet/loss.py
- Run RetinaNet with the configurations in "configs/dr_retina".
Models:
| model | lr sched | multi-scale training | mAP(minival) | mAP (test-dev) | link |
|---|---|---|---|---|---|
| Dr_Retina_R-50-FPN | 1x | No | 37.4 | 37.6 | Google Drive |
| Dr_Retina_R-101-FPN | 2x | Yes | 41.5 | 41.7 | Google Drive |
Citation
If you use the package in your research, please cite our paper:
@inproceedings{qian2020dr,
author = {Qi Qian and
Lei Chen and
Hao Li and
Rong Jin},
title = {DR Loss: Improving Object Detection by Distributional Ranking},
booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR} 2020},
year = {2020}
}