A Contrastive-Learning Framework for Unsupervised Salient Object Detection

This is the official repo for journal paper "A Contrastive-Learning Framework for Unsupervised Salient Object Detection" (TIP2025).

Authors: Huankang Guan, Jiaying Lin, Rynson W.H. Lau

City University of Hong Kong

Installation

We use uv to manage our environment.

uv venv
source .venv/bin/activate
uv pip install torch torchvision numpy pillow progressbar thop pandas opencv-python tqdm joblib albumentations tensorboard
uv pip install git+https://github.com/lucasb-eyer/pydensecrf.git

Or uv pip install -r requirements.txt and then uv pip install git+https://github.com/lucasb-eyer/pydensecrf.git.

Dataset & Checkpoints & Our Results

We zip (dataset, checkpoints and our results) into assets.zip. Download and unzip it to current directory.

Training

source .venv/bin/activate
python train.py --name training_csnet

The output is located in assets/output/training_csnet. Please note that our codebase does not include the self-training stage. If you wish to perform self-training, you can use the output from our CSNet as pseudo labels to train for the final output.

Inference

source .venv/bin/activate
python testmodel.py --weights assets/model_cards/csnet_init_13200.pth --name inference_csnet_init_13200

The output is located in assets/output/inference_csnet_init_13200. If you want results without CRF, run python testmodel.py --weights assets/model_cards/csnet_init_13200.pth --name inference_csnet_init_13200 --crf_round int:0 (faster, No CRF).

Model Complexity

python flops.py --name flops

Our Results

config.json

crf: The number of CRF iteration during training step. [only valid for training]
crf_round: The number of CRF iteration for evaluation. default: 10. If it is set to 0, no CRF is applied. If less than 0, apply bilateral solver.
weights: Full checkpoint.
backboneWeight: backbone checkpoint.

Cite

@ARTICLE{10964591,
  author={Guan, Huankang and Lin, Jiaying and Lau, Rynson W. H.},
  journal={IEEE Transactions on Image Processing}, 
  title={A Contrastive-Learning Framework for Unsupervised Salient Object Detection}, 
  year={2025},
  volume={34},
  pages={2487-2498},
  doi={10.1109/TIP.2025.3558674},
  ISSN={1941-0042},
  month={4}
}