Giters

Pointcept / Pointcept

Pointcept: a codebase for point cloud perception research. Latest works: PTv3 (CVPR'24 Oral), PPT (CVPR'24), OA-CNNs (CVPR'24), MSC (CVPR'23)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About the performance of PTv3 on ScanNet20.

jetpackfirstme opened this issue · comments

commented

Thanks for your great effort in 3D scene understanding! I have a question about PTv3 on ScanNet20. I noticed that for training semantic segmentation on ScanNet20, two losses are used for optimization, namely CrossEntropyLoss, and LovaszLoss. I trained PTv3 with only CrossEntropyLoss and without "normal" feature, and it seems that the performance drops a lot. I wonder have you tried training PTv3 without LovaszLoss? Will it make a big difference and is the phenomenon I encountered a normal one? Looking forward to your reply! Many thanks to you!

commented

Hi, "normal" was introduced since PTv2, and "LovaszLoss" was introduced since PPT. We found that both of them are beneficial to scene understanding. Consequently, we enabled them by default in PTv3 and did not try to disable them.