search

duanyiqun / DiffusionDepth

PyTorch Implementation of introducing diffusion approach to 3D depth perception ECCV 2024
https://arxiv.org/abs/2303.05021
Apache License 2.0
306 stars 17 forks source link

Question about ddim loss #16

Closed wyddmw closed 1 year ago

wyddmw commented 1 year ago

Hi, thanks for the awesome work! When I was reading the code, I noticed there are two implementations for computing the ddim loss which are ddim_loss and ddim_loss_gt, respectively. The primary difference is the choice of x_0. I would like to know if there is an obvious performance gap between these two choices and why using the latent predicted depth output is better. Thanks for the codes again!

duanyiqun commented 1 year ago

Hi there, ddim_loss_gt denotes direct diffusion on GT depth. On the KITTI dataset, it may cause severe overfitting limited to our experiments. If you have different observations we may update them here.

wyddmw commented 1 year ago

Thanks for your reply and this important conclusion. I'll close this issue.