search

guochengqian / PointNeXt

[NeurIPS'22] PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies
https://guochengqian.github.io/PointNeXt/
MIT License
770 stars 111 forks source link

Quick question about PointNeXt #81

Closed SimingYan closed 1 year ago

SimingYan commented 1 year ago

Hi guocheng,

Thanks for your great work! The codebase is also clean and easy to use. May I ask have you tried to use PointNeXt as an encoder backbone for 3d object detection task? According to my understanding, it should also observe the improvement.

guochengqian commented 1 year ago

no. i never worked on detection.
there is another issue asks for the same application. maybe you guys can have a talk.

https://github.com/guochengqian/PointNeXt/issues/78

SimingYan commented 1 year ago

I see. Thanks for your reply! May I ask a quick question about S3DIS segmentation here? I notice you released a trained weight with 71.1 test mIoU but the reported number in the paper is 70.5 $\pm$ 0.3. Is that because this task is unstable and 71.1 is hard to reproduce?

yuchenlichuck commented 1 year ago

Thank you for your questions. For PointNeXt in S3DIS Area 5, the average results without voting in three random runs are reported.