pvrcnn result in pa-aug paper lower than original paper

[turboxin]

Hi @sky77764，thanks for sharing the code.

I have a question when reading the paper, moderate car val AP 3D@0.7 of pvrcnn is 84.84 in original pvrcnn paper and 83.61 in open-pcdet implementation, while in pa-aug paper is 80.43.

Why is it much lower than original paper or open-pcdet implementation.

[sky77764]

@turboxin Sorry for the confusing results.
We reported AP with 11 recall points but pvrcnn with 40.
The AP with 11 recall points is always lower than 40's.

[turboxin]

Hi @sky77764, thanks for the quick reply!

I noticed that pvrcnn also report 11 recall points result which is 83.90

I'm wondering how do you get the 80.43 results since I couldn't find this in pvrcnn paper.

[turboxin]

Sorry the figure in last reply wasn't uploaded successfuly.

[sky77764]

Ah, it's right. I missed it.

These are the raw scores when I report results on the paper.
I only changed the batch size of the original OpenPCDet settings.