search

chenhsuanlin / spatial-transformer-GAN

ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing :eyeglasses: (CVPR 2018)
MIT License
334 stars 72 forks source link

about perturbation #32

Closed shixiao1997 closed 4 years ago

shixiao1997 commented 4 years ago

hello， this great job inspired me a lot, but i have a question about the code. Why add perturbation during training? When I implement stn by myself, it also work normally without him.

chenhsuanlin commented 4 years ago

This is because we want to simulate different incorrect configurations that the network might see. If you could get good results without data augmentation, it probably means the training data itself cover sufficient variations already.