The result of PointPillars Anchor-based and Center-based is different.

[Geunju-hub]

@tianweiy
I am currently working on improving the performance of Point Pillars' Kitti 3d data set.
In your paper, when the head of PointPillars was center-based, you mentioned that the performance improvement was made in nuScenes dataset.
So I checked whether the performance improvement was made in kitti dataset as well, but the performance was rather poor.
Could you give me an additional explanation on this part?

I conducted an experiment using the mmdetection3d library.

I used for detect Car, Pedestrian, Cyclist.

the config only different of head.

And the results are below.

---

Overall 3D AP@40	Easy	Moderate	Hard
PointPillars	76.5073	64.7261	61.3703
PointPillars + CenterPointHead	72.5513	61.0547	58.1883

[thatnn]

Can u provide PointPillats +CenterPointHead yaml & ckpt?

email: thatnn@ajou.ac.kr

[Geunju-hub]

Hi My name is Geunju Baek. from Sejong University, Korea. First of all, Thank you for giving me a question. Actually, I've finished my experiment 3 months ago. Here is my config file work on mmdet3d. And my checkpoint is below. https://drive.google.com/file/d/1-oJ0zOzgoKC9yhGmo44t8iLo5qSs2YEV/view?usp=sharing Finally, u can see my paper review about 3d object detection using with point cloud. Geunju-hub/3D-detection-Paper-Review: Sejong creative semester system - 3D detection Paper Review (github.com) <https://github.com/Geunju-hub/3D-detection-Paper-Review/tree/main> Best regards Geunju Baek 2023년 8월 29일 (화) 오전 11:25, thatnn ***@***.***>님이 작성:

…

Can u provide PointPillats +CenterPointHead yaml & ckpt? email: ***@***.*** — Reply to this email directly, view it on GitHub <#50 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQJKSX7XG32CX7NEYVLLIITXXVHIFANCNFSM6AAAAAAXZH6ZIU> . You are receiving this because you authored the thread.Message ID: ***@***.***>