You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This is a question about the method/paper, not so much the implementation.
During training, do you provide the corresponding camera pose (denoted as ξ in the paper) to the discriminator? It appears the answer is no. If this is the case, why doesn't the generator just ignore the camera pose altogether, and just learn to generate images from a random angle each time? In my mind, the discriminator wouldn't be able to tell? Perhaps you train on multiple samples with the same z per batch, enforcing that different ξ give reasonable results for the same z?
The text was updated successfully, but these errors were encountered:
This is a question about the method/paper, not so much the implementation.
During training, do you provide the corresponding camera pose (denoted as ξ in the paper) to the discriminator? It appears the answer is no. If this is the case, why doesn't the generator just ignore the camera pose altogether, and just learn to generate images from a random angle each time? In my mind, the discriminator wouldn't be able to tell? Perhaps you train on multiple samples with the same z per batch, enforcing that different ξ give reasonable results for the same z?
The text was updated successfully, but these errors were encountered: