DGDet is a method for improving performance in an out-of-distribution (OOD) environment while maintaining detection performance in an in-distribution environment.
This code is based on MMdetection open source.
This project is released under the Apache 2.0 license.
Please refer to get_started.md in MMdetection for installation.
Please refer to get_started.md for the basic usage of MMDetection.
- Linux or macOS (Windows is in experimental support)
- Python 3.6+
- PyTorch 1.3+
- CUDA 9.2+ (If you build PyTorch from source, CUDA 9.0 is also compatible)
- GCC 5+
- MMCV
- WandB
- Pillow 7.2.0
We also provide docker files for building the environment.
docker pull dshong/mmdetection:1.6.0-cuda10.1-cudnn7docker pull dshong/mmdetection:1.8.0-cuda10.1-cudnn7docker pull dshong/mmdetection:1.10.2-cuda10.1-cudnn7docker pull dshong/mmdetection:1.11.0-cuda11.3
NOTE: The docker file above does not contain WandB.
To use WandB, please complete the installation through pip install wandb and login through wandb login.
The running settings are included in the configuration file.
The configuration files are located in the configs folder.
You can train your model by passing keyword arguments to train.py as shown below.
-
base (8epochs)
python3 /ws/external/tools/train.py /ws/external/configs/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py --work-dir /ws/data/ai28/cityscapes/faster_rcnn_r50_fpn_1x_cityscapes
-
w/ AugMix (2epochs)
python3 /ws/external/tools/train.py /ws/external/configs/cityscapes/2epoch/faster_rcnn_r50_fpn_1x_cityscapes_augmix.py --work-dir /ws/data/ai28/cityscapes/2epoch/faster_rcnn_r50_fpn_1x_cityscapes_augmix
-
DGDet (w/ AugMix) (2epochs)
python3 /ws/external/tools/train.py /ws/external/configs/cityscapes/2epoch/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100.py --work-dir /ws/data/ai28/cityscapes/2epoch/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100
The model can be tested in the same way as below.
-
base
python3 /ws/external/tools/test.py /ws/data/ai28/faster_rcnn_r50_fpn_1x_cityscapes/faster_rcnn_r50_fpn_1x_cityscapes.py /ws/data/ai28/faster_rcnn_r50_fpn_1x_cityscapes/epoch_8.pth --work-dir /ws/external/ai28/faster_rcnn_r50_fpn_1x_cityscapes --eval bbox
-
DGDet (w/ AugMix) (2epochs)
python3 /ws/external/tools/test.py /ws/data/ai28/2epoch/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100.py /ws/data/ai28/2epoch/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100/epoch_2.pth --work-dir /ws/external/ai28/2epoch/augmix.wotrans_plus_rpn.jsdv1.3.none_roi.jsdv1.3.none__e2_lw.1e-1.100 --eval bbox
NOTE: Before running, make sure you have chosen the correct configuration file.
Results on CityScapes with Faster-RCNN:
| Detector | Arch | RPN Loss | RoI Loss | Lambda weight | Clean mAP(%) | Corruption mPC(%) | |
|---|---|---|---|---|---|---|---|
| Base | Faster-RCNN | ResNet50 | CrossEntropy + L1loss | CrossEntropy + SmoothL1loss | - | 40.6 | 11.0 |
| w/ AugMix | Faster-RCNN | ResNet50 | CrossEntropy + L1loss | CrossEntropy + SmoothL1loss | - | 42.8 | 16.0 |
| DGDet | Faster-RCNN | ResNet50 | CrossEntropyPlus(jsdv1.3) + L1loss | CrossEntropyPlus(jsdv1.3) + SmoothL1loss | 0.1(RPN), 100(RoI) | 40.2 | 20.9 |
| DGDet | Faster-RCNN | ResNet50 | CrossEntropyPlus(jsdv1.3) + L1loss | CrossEntropyPlus(jsdv1.3) + SmoothL1loss | 0.1(RPN), 10(RoI) | 43.0 | 20.0 |
@article{mmdetection,
title = {{MMDetection}: Open MMLab Detection Toolbox and Benchmark},
author = {Chen, Kai and Wang, Jiaqi and Pang, Jiangmiao and Cao, Yuhang and
Xiong, Yu and Li, Xiaoxiao and Sun, Shuyang and Feng, Wansen and
Liu, Ziwei and Xu, Jiarui and Zhang, Zheng and Cheng, Dazhi and
Zhu, Chenchen and Cheng, Tianheng and Zhao, Qijie and Li, Buyu and
Lu, Xin and Zhu, Rui and Wu, Yue and Dai, Jifeng and Wang, Jingdong
and Shi, Jianping and Ouyang, Wanli and Loy, Chen Change and Lin, Dahua},
journal= {arXiv preprint arXiv:1906.07155},
year={2019}
}