The smoothloss - Githubissues

hisfog / SfMNeXt-Impl

[AAAI 2024] Official implementation of "SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation", and more.

MIT License

85 stars 12 forks source link

The smoothloss #5

Closed Jamesgender closed 1 year ago

Jamesgender commented 1 year ago

Hello! My question is about smoothloss. This version code shows that the output of the model is depth. And you still use 'outputs["disp", 0] = pred' to save the depth you predicted. That is ok. So you changed the behavior in generate_images_pred function. But in smoothloss, it seems that you still use depth to compute the smoothloss. Why didnt you use the inverse depth, which means the true disparities for smoothloss as the monodepth2 did? The results will be the same? Thanks!

hisfog commented 1 year ago

Yes, we calculate the smooth loss using depth, and it is feasible.
It should also be feasible to calculate the smooth loss using 1 / depth (and it seems more reasonable).
It seems that there is currently no paper indicating which one is superior. I think we should conduct more experiments to investigate which one is better.

Jamesgender commented 1 year ago

OK，thank you for your timely answer！