This code was implemented with Python 3.6, PyTorch 1.8.0, and CUDA 11.1 on an NVIDIA GeForce GTX 3090.

!!!2023-12-4. We have updated our saliency maps. We sincerely apologize for a previous mistake in reading images during the generation of general visual representations on the test dataset, leading to inaccurate results on the DUT-OMRON dataset in the paper. We have released the revised saliency maps. Overall, our method is still effective and remains significantly leading in weakly-supervised SOD and approaching fully-supervised methods.

Download the DUTS and other datasets and unzip them into VRF/data folder. Scribble labels can be downloaded from Scribble Saliency.

Clone the Tokencut project

git clone https://github.com/YangtaoWANG95/TokenCut

Enter the directory

cd Tokencut

Install dino_deitsmall8_pretrain from https://dl.fbaipublicfiles.com/dino/dino_deitsmall8_pretrain/dino_deitsmall8_pretrain.pth. Then, copy our provided main_token_similarity_save.py into this directory.

General visual representation generation

python main_token_similarity_save.py --dataset 'DUTS' --mode 'train'

The general visual representations of the test dataset are also generated in this way, such as the DUTS-TE dataset.

python main_token_similarity_save.py --dataset 'DUTS-TE' --mode 'test'

cd VRF

Download the pretrianed weight for backbone Res50.

#Train
python train_token_similarity_40_40_384_mul_dino.py

Pretrained models can be downloaded from Google Drive. !!!The general visual representations of the test dataset should be generated before starting testing. (refer to Part 2)

#Prediction
python test_dino.py --model_path path/to/pretrained_model/folder/

We provide the pre-computed saliency maps Google Drive.

Thanks to Tokencut and SCWSSOD.