FPS method issue in your code

[FrankCAN]

Hi, thanks for your impressive work. I noticed in your code that you use "FS" method for FPS downsampling in Layer 2, but why did you separate and use "F-FPS, D-FPS" in Layer 3? Is there any difference for both methods?

When you use "FS" in Layer 2, you downsample 4096 points to 512 points separately by F-FPS and D-FPS, and then concatenate them together. But when you use "F-FPS, D-FPS" in Layer 3, you firstly split 1024 points to two parts, each of which has 512 points. And then you respectively use F-FPS and D-FPS to downsample corresponding 512 points to 256 points, and finally concatenate to 512 points.

I am confused that why you didn't just use "FS" directly?

Thanks for your help.

Best Regards Frank

[lozino]

I guess that doing it that way allows them to sample with F-FPS only the set of points that was previously selected with F-FPS and with D-FPS only the set of points that was previously selected with D-FPS, instead of mixing the two sets together.

[kaiopen]

It is my understanding that "FS" is different from the combination of "F-FPS" and "D-FPS". Actually "FS" does "F-FPS" and "D-FPS" separately. Here, Both "F-FPS" and "D-FPS" sample points from all 4096 points. So we get 1024 points where 512 points from "F-FPS" picking up from all 4096 points and 512 points from "D-FPS" picking up from all 4096 points. Note that in 1024 points, the top 512 points are from "F-FPS" and the last 512 points are from "D-FPS". In Layer 3, do "F-FPS" on the top 512 points from the previous "F-FPS" for 256 points and do "D-FPS" on the last 512 points from the previous "D-FPS" for 256 points. And I agree with @lozino .In my opinions, "F-FPS" is better for picking up foreground points as it is said in the paper. As keeping doing "F-FPS" on the points from "F-FPS", less background points will be got. In other words, foreground points will be got. And in vote layer, offsets for foreground points are expected, not background points.