search

nv-tlabs / semanticGAN_code

Official repo for SemanticGAN https://nv-tlabs.github.io/semanticGAN/
MIT License
180 stars 34 forks source link

Generating image label pairs with dataset containing more labeled than unlabeled images #34

Open pcicales opened 2 years ago

pcicales commented 2 years ago

Hello

I am working with datasets that contain more labeled images than unlabeled. In my experiments with default settings, it seems that the generator struggles to learn the input, and images can be quite noisy. Is there a reason for this? From what I can tell, more labels should not harm the generator but I may be wrong.

This is for single class images with very little background, as they are cropped to the height and width of the desired object.

lidaiqing commented 2 years ago

Hi @pcicales , thank you for your interests in our work. Could you share an example of your image label pairs? With more labelled data, it should learn a better joint distribution. It is likely the data processing step is wrong.