Are camera geometry tensor and self-attention autoencoder included
Opened this issue · 0 comments
bolianchen commented
Dear Authors
Thanks for sharing such excellent work.
I have roughly checked the codes and could not find them.
From the compute_losses method in PhotometricReconstructionLoss, it seems only reconstruction loss and smoothness loss are applied. In the PoseDecoder, only the last feature from the encoder is considered without any shortcuts. Not to mention the camera geometry tensor which is involved in the shortcuts as mentioned in the paper.
Thanks and appreciate any feedback.
Best,
Bolian