search

hansen7 / OcCo

[ICCV '21] "Unsupervised Point Cloud Pre-training via Occlusion Completion"
https://hansen7.github.io/OcCo/
MIT License
233 stars 26 forks source link

the log of train_completion.py #30

Open zshyang opened 2 years ago

zshyang commented 2 years ago

Hi @hansen7 , Thanks for sharing this great work with us. I have a question regarding the output log during pretraining. The detail is here. The question is, it seems that the loss did not drop too much or its range is varied from 0.03 to 0.04. Did this happen to your pretraining as well? But visually, I could see some improvements in the plots folder. Best regards

hansen7 commented 3 months ago

Thank you, @zshyang. In my experience, the pretraining loss also reached saturation quickly.

Regarding the scenario where better visualizations with a slightly increased loss, I believe this is due to the Chamber Distance loss being poorly calibrated. It does not accurately represent the true quality of the generated point clouds.