search

yaoqi-zd / SGAN

Saliency Guided Self-attention Network for Weakly and Semi-supervised Semantic Segmentation(IEEE ACCESS)
MIT License
46 stars 12 forks source link

Training using my own dataset #15

Closed gruossomonica closed 3 years ago

gruossomonica commented 3 years ago

Hi, thanks for your interesting work. Can I perform training using your approach with my own dataset? I collected about 5k labeled images. What are the tips and steps in this case? I look forward to your reply. Thank you.

Best Regards,

Monica

yaoqi-zd commented 3 years ago

@gruossomonica Hi, if your collected 5k images are labeled with segmentation mask, I suggest you using fully-supervised or semi-supervised methods to fully exploit the label information. If your labels are image-level tags, then you can follow the steps: 1) use a pretrained saliency model to extract saliency mask for each labeled images, such as SNet, DHS in our paper; 2) train a classifier on the labeled images and generate initial seeds with high thresholds; 3) use the saliency mask, initial seeds and original images to train SGAN, and generate more complete seeds; 4) use the seeds from 3) to train a segmentation model in fully-supervised manner the detailed description and commands can be found in the README

gruossomonica commented 3 years ago

Thank you for the support. Best Regards, Monica