Open SPA-junghokim opened 3 months ago
For fair comparison, we follow loss setting in RenderOcc, while the Rendering Supervision with 2D Labels in renderocc is very time-consuming. The training time would be same with flashocc when the Rendering Supervision with 2D Labels is discard. And we found that the Rendering Supervision with 2D Labels makes little contribution to the performance, the improvement of renderocc to flashocc mainly comes from the class-balance weight.
Thank you for sharing your great work.
I have a question about training time.
In my experiments, it seemed to require 7 days more for 24 epoch with 4 3090 RTX gpus. Is flashocc-render model required 7 times more than Original Flashocc?? If then, why does render model require more times??
Thx. Jungho Kim.