Improving speed of Pointpillars

[xavidzo]

Hello @AutoVision-cloud, thanks for the release of your nice work.

I have a question regarding the speed of Pointpillars. Since you determined that Pointpillars-FSA (or -DSA) runs with nearly half the amount of G-FLOPs, I assume this means your model should have a faster inference speed than the baseline...
if my assumption is correct, do you know how much faster can it run?
I tested Pointpillars-FSA with custom data and compared the speed with base Pointpillars, but didn't see any difference in speed....
To me it's not clear how Pointpillars-FSA would perform less G-FLOPs if you are adding extra layers to the baseline, con you please explain? Or for faster speed should I pass only the context features to the BEVEncoder, and ignore the pillar features, thus avoiding concatenation of pillar and context features?