search

ZM-Zhou / SDFA-Net_pytorch

Apache License 2.0
17 stars 4 forks source link

Regarding image reconstruction #8

Open seongwn opened 1 year ago

seongwn commented 1 year ago

Thanks for your work

I have a question regarding image reconstruction.

In stage 1, you reconstruct right image and in stage2 you reconstruct left image to calculate Photometric loss. Is there a specific reason for this selection?

Thank you.

ZM-Zhou commented 1 year ago

Hi, Thanks for your attention! For learning the left-view depth map, we reconstruct the right-view image for training in both stage1 and stage2, which we called image synthesis loss. I think that the photometric loss that you have mentioned is used to calculate the photometric principle mask, which is used to select the reliable depths on the left-view depth map for self-distillation.