Home

Awesome

UMNet

The Pytorch implementation of CVPR2022 paper Multi-Source Uncertainty Mining for Deep Unsupervised Saliency Detection

Trained Model,Test Data and Results

Please download the trained model, test data and SOD results from Baidu Cloud (password: tmzw).

Requirement

• Python 3.7

• PyTorch 1.6.1

• torchvision

• numpy

• Pillow

• Cython

Run

  1. Please download the trained model and test datasets (including DUTS-TE, OMRON, ECSSD, and HKU-IS). Uncompress and put them in the current file.
  2. Set the path of testing sets and trained model in config.py. The default setting can be in config.py.
  3. Run main.py to obtain the predicted saliency maps. The results are saved in the save_path (see config.py). You can also download our saliency results from Baidu Cloud.
  4. Run compute_score.py to obtain the evaluation scores of the predictions in terms of MAE, Fmax, Sm, and Em. The evaluation codes are referred from https://github.com/Xiaoqi-Zhao-DLUT/GateNet-RGB-Saliency.
  5. Please be sure that the paths of ground truth and predictions are valid in compute_score.py.

Train

Note: Our method is trained mainly following the same setting of DeepUSPS. We use MSRA-B 2500 training data for network training.

  1. Four traditional SOD methods including MC, HS, DSR, and RBD are adopted to generate pseudo labels for the training data, which are refined using the first stage of DeepUSPS.
  2. The four kinds of refined pseudo labels are used for multi-source network learning using our training code (extract code: a4hh).