Max F-measure: 91.75% [Baseline: 90.61%]

1. Download the pretrained model [Google Drive|BaiduYunPan].
2. Try it now!

   python predict.py --im_path [IM_PATH] \
                     --netG [MODEL_PATH] \
                     --thres [-1|161] \
                     --dgf --nn_dgf \
                     --post_sigmoid --cuda

1. Download and unzip the saliency dataset MSRA-B.

2. Make A-B (Image-Label) pairs.

   python scripts/preprocess.py --data_path [MSRA-B_ROOT] --mode train
   python scripts/preprocess.py --data_path [MSRA-B_ROOT] --mode valid
   python scripts/preprocess.py --data_path [MSRA-B_ROOT] --mode test

3. Start training!

   • WITHOUT Guided Filtering Layer

   python main.py --dataroot [MSRA-B_ROOT]/AB --cuda --experiment [EXP_NAME]
   • WITH Guided Filtering Layer

   python main.py --dataroot [MSRA-B_ROOT]/AB --cuda --experiment [EXP_NAME] --dgf
   • Finetune

   python main.py --dataroot [MSRA-B_ROOT]/AB --cuda --experiment [EXP_NAME] --netG [MODEL_PATH] --dgf

4. Evaluation

   python test.py --dataroot [MSRA-B_ROOT]/AB/test \
                  --netG [MODEL_PATH] --cuda \
                  --experiment [SAVE_FOLDER] \
                  --nn_dgf --post_sigmoid --dgf

5. Calculate metrics

   1. Install SalMetric.

      git clone https://github.com/Andrew-Qibin/SalMetric && cd SalMetric 
      mkdir build && cd build
      cmake .. && make
   2. Calculate !

      cd [SAVE_FOLDER]
      [SalMetric_ROOT]/build/salmetric test.txt [WORKER_NUM]

A part of the code was adapted from DSS.