code for "Learning to Detect Salient Object with Multi-source Weak Supervision"
The SNet is supervised by the natural image dataset with noisy pseudo ground-truth predicted by CNet and PNet on unlabeled saliency images and a synthetic image dataset created by composing detected salient objects and background images.
- To generate the saliency masks of the above two dataset, run
to be added XXXXX
The averaged map is processed with CRF and then binarized.
- To create the synthetic image dataset, run
python test_crf.py
python syn_msk.py
- train SNet, run
python train_sal.py --name save_folder_name --model SalModel --batchSize 24
- test the results of the SNet, run
python test_sal.py --name save_folder_name --model SalModel --batchSize 24