search

autonomousvision / differentiable_volumetric_rendering

This repository contains the code for the CVPR 2020 paper "Differentiable Volumetric Rendering: Learning Implicit 3D Representations without 3D Supervision"
http://www.cvlibs.net/publications/Niemeyer2020CVPR.pdf
MIT License
798 stars 90 forks source link

clear on depth back propagation #5

Closed hua-wu-que closed 4 years ago

hua-wu-que commented 4 years ago

Hi! I try to understand these words from section 3.3 but get very confused: "In practice, however, we would like to implement the backward pass not only for a single surface depth ˆd but for a whole batch of depth values."

Could you please explain why the backward pass not tackle for a single depth value but for multiple?

Thank you!

m-niemeyer commented 4 years ago

Hi @hua-wu-que , thanks for your interest!

Regarding your question, what we mean here is that in a single iteration, we do not only sample a single pixel but a batch of pixels (1024) which we process in parallel. So during the backward pass, we need to handle multiple depth predictions in parallel, where each corresponds to a single pixel. In principle, you are right, we could train with only a single pixel per iteration (and hence a single depth value in the backward pass), but training would be very slow and also very stochastic.

I hope this helps. Good luck with your research!

hua-wu-que commented 4 years ago

Thank you for the reply! closing this issue.