Synthesize Boundaries: A Boundary-aware Self-consistent Framework for Weakly Supervised Salient Object Detection

Usage

1. Download

1.1 Download the DUTS and other datasets and unzip them into framework/data folder. Scribble labels can be downloaed from Scribble Saliency.

1.2 Download ResNet-50 pretrained models and save it into framework folder (Res50)

2. Synthetic Image Generation

python synthetic_image_generation/main.py

save it into framework/data folder

3. Train 

python framework/train.py

4. Predict

Our saliency detector is SCWSSOD, so the saliency maps can be generated by combining the trained model parameters with the prediction code of SCWSSOD. Please refer to SCWSSOD.

Pretrained models can be downloaded from Google Drive.

We provide the pre-computed saliency maps from our paper Google Drive.

Thanks to SCWSSOD.