The input images to train

[smmdream]

hello chick: I find when I use the cityScapes dataset,I need to convert to get gtFine_labelTrainIds.png, the gtFine_labelTrainIds.png is created by _gt_polygons.json. now,I want to use PSPNet on other dataset,I have a question ,when I train the network,what is the format of gt_file_image,are these such as gtFine_labelTrainIds.png? or the color.png is OK?

[hellochick]

Hey @smmdream , I think you should preprocess the color.png first, and let the pixel value range from 0~num_classes, just like gtFine_labelTrainIds.png. If you use RGB value to train, it cannot converge.

[smmdream]

Thank you, @hellochick I'll try