NVlabs/ssn_superpixels

custom training

Zampal94 opened this issue · 1 comments

Hello, thanks for sharing the code. Does it make sense to train the model with a dataset of images with varying size and resolution or would it never converge?

SSN might be resolution-specific in its use of CNN features which are usually resolution-specific to some extent. You can definitely try this if you use a CNN base that is robust to resolution changes.