Single-DGOD
Single-DGOD copied to clipboard
Single-Domain Generalized Object Detection
CVPR2022: Single-Domain Generalized Object Detection in Urban Scene via Cyclic-Disentangled Self-Distillation.
The current code is Faster R-CNN with FPN. In our paper, we do not utilize FPN.

Datasets
Daytime-Sunny, Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy
[models]
Training
CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net_fpn.py
--dataset dc_fpn --net res101 --epochs 20
--bs 2 --nw 8
--lr 0.004 --lr_decay_step 8
--cuda
Evaluation
CUDA_VISIBLE_DEVICES=$GPU_ID python test_net_fpn.py --dataset dc_fpn --dataset_test voc_2007_train_nightclear --net res101
--checksession 1 --checkepoch 10 --checkpoint 19317
--cuda
New Results

Citation
If you find this repository useful for your work, please cite as follows:
@inproceedings{wu2022single,
title={Single-Domain Generalized Object Detection in Urban Scene via Cyclic-Disentangled Self-Distillation},
author={Wu, Aming and Deng, Cheng},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={847--856},
year={2022}
}