/MIRV

Mutual Information Regularization for Weakly-supervised RGB-D Salient Object Detection

Primary LanguagePython

Mutual Information Regularization for Weakly-supervised RGB-D Salient Object Detection (TCSVT2023)


[paper_link]

Set up

  • pip install -r requirements.txt
  • cd ./first_stage/kernels/lib_tree_filter
  • python setup.py build develop

Train Model

  • Prepare data for training the first stage (We provided the related data in:train_data. Please download it and put it in the '/train_data/' folder, and download the VGG model and put it in './first_stage/vgg_models/': vgg_pretrain_model)
  • Run ./first_stage/train.py
  • Prepare data for training the MVAE refine (We provided the pseodu label of train_data tested by the first_stage model in:MVAE_refine. Please download it and put it in the "./first_stage/models/results_train/" folder)
  • Run ./MVAE_refine/train.py

Test Model

  • Run ./first_stage/test.py
  • Run ./MVAE_refine/test.py

Trained model:

Please download the trained model and put it in "./first_stage/models": [Google_Drive];
"./MVAE_refine/models": [Google_Drive]

Prediction Maps (Permissions are open)

Results of our model on six benchmark datasets(NJU2K, SSB, DES, NLPR, LFSD, SIP) can be found: first_stage;MVAE_refine