CamoDiffusion

This repository is the official implementation of CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models.

Our implementation is based on the denoising diffusion repository from lucidrains, which is a PyTorch implementation of DDPM.

And we provide our pretrained weight and inference result in release.

Requirements

• python == 3.9
• cuda == 11.3

To install requirements:

pip install -r requirements.txt

Dataset

COD (Camouflaged Object Detection) Dataset

Training

To train the model(s) in the paper, run this command:

accelerate launch train.py --config config/camoDiffusion_352x352.yaml --num_epoch=150 --batch_size=32 --gradient_accumulate_every=1

And then finetune it to 384 size:

accelerate launch train.py --config config/camoDiffusion_384x384.yaml --num_epoch=20 --batch_size=28 --gradient_accumulate_every=1 --pretrained model_352/model-best.pt --lr_min=0 --set optimizer.params.lr=1e-5

Evaluation

To test a model, run sample.py with the desired model on different datasets:

accelerate launch sample.py \
  --config config/camoDiffusion_384x384.yaml \
  --results_folder ${RESULT_SAVE_PATH} \
  --checkpoint ${CHECKPOINT_PATH} \
  --num_sample_steps 10 \
  --target_dataset CAMO \
  --time_ensemble

For ease of use, we create a eval.sh script and a use case in the form of a shell script eval.sh. You can edit the script to change the parameters you want to test.

bash scripts/eval.sh

Citation

@article{chen2023camodiffusion,
  title={CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models},
  author={Chen, Zhongxi and Sun, Ke and Lin, Xianming and Ji, Rongrong},
  journal={arXiv preprint arXiv:2305.17932},
  year={2023}
}