Closed imbinwang closed 3 years ago
This is because we first trained the model with correspondence loss with “False" flag for spatio-aware module. Then we trained the spatio-aware module without this loss.
Thank for your clarification. As I understand, you means the training process consist of two steps:
Do you train the whole model in a process? It seems a valid choice.
The correspondence loss is used to bridge the gap between the source image and the target image in different poses. This is for computing the correspondence of these two features. You can check cocosnet. Though these two features do not have large gaps, but we use it to obtain a more precise correspondences. In fact, if do not consider the training time, you can train using spatio-aware module with the correspondence loss.
I will try it. Thank you.
Hi,
The correspondence loss of image generator in paper section 3.3 is not used in your code.
Does the correspondence loss not impact the final generated image? Thanks for your reply.