Awesome

LPLD (Low-confidence Pseudo Label Distillation) (ECCV 2024)

We are currently working on refactoring all the original code. Please wait for the final version. However, you can run the example code using the instructions below.

This is an official code implementation repository for Enhancing Source-Free Domain Adaptive Object Detection with Low-confidence Pseudo Label Distillation, accepted to ECCV 2024.

Installation and Environmental settings (Instructions)

• We use Python 3.6 and Pytorch 1.9.0
• The codebase from Detectron2.

git clone https://github.com/junia3/LPLD.git

conda create -n LPLD python=3.6
conda activate LPLD
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cudatoolkit=10.2 -c pytorch

cd LPLD
pip install -r requirements.txt

## Make sure you have GCC and G++ version <=8.0
cd ..
python -m pip install -e LPLD

Dataset preparation

• Cityscapes, FoggyCityscapes / Download Webpage / Google drive (preprocessed)
• PASCAL_VOC / Download Webpage
• Clipart / Download Webpage / Google drive (preprocessed)
• Watercolor / Download Webpage / Google drive (preprocessed)
• Sim10k / Download Webpage

Make sure that all downloaded datasets are located in the ./dataset folder. After preparing the datasets, you will have the following file structure:

LPLD
...
├── dataset
│   └── foggy
│   └── cityscape
│   └── clipart
│   └── watercolor
...

Make sure that all dataset fit the format of PASCAL_VOC. For example, the dataset foggy is stored as follows:

$ cd ./dataset/foggy/VOC2007/
$ ls
Annotations  ImageSets  JPEGImages
$ cat ImageSets/Main/test_t.txt
target_munster_000157_000019_leftImg8bit_foggy_beta_0.02
target_munster_000124_000019_leftImg8bit_foggy_beta_0.02
target_munster_000110_000019_leftImg8bit_foggy_beta_0.02
.
.

Execution

Currently, we only provide code and results with ResNet-50 backbone baselines. We are planning to add VGG-16 backbone baselines and code.

Test models

CUDA_VISIBLE_DEVICES=$GPU_ID python tools/test_main.py --eval-only \ 
--config-file configs/sfda/sfda_city2foggy.yaml --model-dir $WEIGHT_LOCATION

Visualize

We provide visualization code. We use our trained model to detect foggy cityscapes in the example image.

Pretrained weights (LPLD)

Citation

@article{yoon2024enhancing,
  title={Enhancing Source-Free Domain Adaptive Object Detection with Low-confidence Pseudo Label Distillation},
  author={Yoon, Ilhoon and Kwon, Hyeongjun and Kim, Jin and Park, Junyoung and Jang, Hyunsung and Sohn, Kwanghoon},
  journal={arXiv preprint arXiv:2407.13524},
  year={2024}
}