search

duanyiqun / DiffusionDepth

PyTorch Implementation of introducing diffusion approach to 3D depth perception
https://arxiv.org/abs/2303.05021
Apache License 2.0
293 stars 16 forks source link

Loss Difference between paper and code #6

Closed RickyYXY closed 1 year ago

RickyYXY commented 1 year ago

I notice that you have mentioned a latent space loss in the paper(equ.7). But I can’t find this loss in the training code. So which one is right? This is strange.

duanyiqun commented 1 year ago

Hi Ricky, we found after the paper's formation that, the latent space loss only boosts slightly higher performance but with a much higher computational cost. So we remove the latent loss in the public version. Regards

RickyYXY commented 1 year ago

Get it! And I find that the depth pixel loss is directly L1 L2 loss, not like loss in Equ.6. Is it because using Equ.6 makes no difference? Thanks for replying me so fast.

duanyiqun commented 1 year ago

Not this case, L1 L2 works well for general. But you may use the original loss in the paper called SIGloss (same as depth former and binsformer) in the loss subfolder, this is to boost the metric performance.

RickyYXY commented 1 year ago

OK, so if I want the best performance, I should change the default config and use the SIGloss instead of L1+L2 loss, right?

duanyiqun commented 1 year ago

I would suggest for SIGLoss + L1+L2 + DDIM loss for initialization and SIGLoss + DDIM for finetune.

RickyYXY commented 1 year ago

OK! Thanks for your reply! Helped me a lot!