Awesome
<p align="center"> <h1 align="center">Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation </h1> <p align="center"> <a href="https://zpdu.github.io/">Zhipeng Du</a> Β· <a href="https://sites.google.com/site/miaojingshi/home">Miaojing Shi</a> Β· <a href="https://jiankangdeng.github.io/">Jiankang Deng</a> </p>PyTorch implementation of Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation. (CVPR 2024) [Page | Paper]
π¨ To-Do List
-
- release the code regarding the proposed model and losses.
-
- release the evaluation code, and the pretrained models.
-
- release the training code.
:rocket: Installation
Begin by cloning the repository and setting up the environment:
git clone https://github.com/ZPDu/DAI-Net.git
cd DAI-Net
conda create -y -n dainet python=3.7
conda activate dainet
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
:notebook_with_decorative_cover: Training
Data and Weight Preparation
- Download the WIDER Face Training & Validation images at WIDER FACE.
- Obtain the annotations of training set and validation set.
- Download the pretrained weight of Retinex Decomposition Net.
- Prepare the pretrained weight of the base network.
Organize the folders as:
.
βββ utils
βββ weights
β βββ decomp.pth
β βββ vgg16_reducedfc.pth
βββ dataset
β βββ wider_face_train.txt
β βββ wider_face_val.txt
β βββ WiderFace
β β βββ WIDER_train
β β βββ WIDER_val
Model Training
To train the model, run
python -m torch.distributed.launch --nproc_per_node=$NUM_OF_GPUS$ train.py
:notebook: Evaluationβ
On Dark Face:
- Download the testing samples from UG2+ Challenge.
- Download the checkpoints: DarkFaceZSDA (28.0) or DarkFaceFS (52.9, finetuned with full supervision).
- Set (1) the paths of testing samples & checkpoint, (2) whether to use a multi-scale strategy, and run test.py.
- Submit the results for benchmarking. (Detailed instructions).
On ExDark:
- Our experiments are based on the codebase of MAET. You only need to replace the checkpoint with ours for evaluation.
π Citation
If you find this work useful, please cite
@inproceedings{du2024boosting,
title={Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation},
author={Du, Zhipeng and Shi, Miaojing and Deng, Jiankang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={12666--12676},
year={2024}
}
or
@article{du2023boosting,
title={Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation},
author={Du, Zhipeng and Shi, Miaojing and Deng, Jiankang},
journal={arXiv preprint arXiv:2312.01220},
year={2023}
}
π Acknowledgement
We thank DSFD.pytorch, RetinexNet_PyTorch, MAET, HLA-Face for their amazing works!