search

ziyc / drivestudio

A 3DGS framework for omni urban scene reconstruction and simulation.
https://ziyc.github.io/omnire/
MIT License
595 stars 51 forks source link

On generating masks with bounding boxes #44

Closed Enamel2 closed 3 weeks ago

Enamel2 commented 1 month ago

Thanks for the excellent and inspiring work! If I may, I would like to ask a little question I encountered when I writing my own preprocess script.

I noticed the dynamic mask generation process will exclude partially visible objects, so that image region of those partially visible dynamic objects will not have a mask on them. Would this pollute the training for the background GS model? Or am I missing something?

Thanks a lot for your time answering my question!

ziyc commented 3 weeks ago

Hi @Enamel2, thanks for your question!

The dynamic masks are only used when evaluating the trained results - we use them to mask dynamic regions and compute metrics specifically for those areas. While the dynamic mask processing may occasionally have prediction errors from the masking models, it's tolerable and it doesn't affect the training process. During training, we only use sky masks for sky supervision.

Enamel2 commented 3 weeks ago

Thank you for the elaborated explanation! This helps a lot!