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. Besides, welcome to focus on our CVPR 2024: Prompt-Driven Dynamic Object-Centric Learning for Single Domain Generalization. [https://github.com/Daniel00008/PDOC]

Datasets

Daytime-Sunny, Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy

[Download link]

[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}
}