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

A question about the pred noise. #15

Open northagain opened 1 year ago

northagain commented 1 year ago

Thanks for ur excellent work! I have a question about ddim loss which is computed with noise and pred noise. The pred noise, which is the output of ScheduledCNNRefine Model is >0, because the last activation function is RELU. However, shouldn't the randn noise contains both >0 parts and <0 parts? image

duanyiqun commented 1 year ago

It is a good question. It should be theoretically right in this case. Have you tried to use LeakyRelu instead for training? Limited to my memory, I didn't remember a clear difference. But I would appreciate it if you have some comparison results between ReLU and LeadkyReLU. If the results are better, I may have an update. Or we may have a pull request together.