search

xinge008 / Cylinder3D

Rank 1st in the leaderboard of SemanticKITTI semantic segmentation (both single-scan and multi-scan) (Nov. 2020) (CVPR2021 Oral)
Apache License 2.0
859 stars 180 forks source link

NuScenes multiple sweeps #74

Closed ShengyuH closed 3 years ago

ShengyuH commented 3 years ago

hi,

Congrats on this good work.

I am recently checking NuScene Lidarseg benchmark, according to its official instructions, we are allowed to used up to 10 past sweeps to density the point clouds and hopefully achieve better perception. However, if I understand your codebase correctly, you only use the key frame and also achieve competitive performance(ranked 4th). Did you try models with accumulated point clouds and what’s your experience with it? I know that on semanticKITTI they distinguish between moving and non-moving counterparts and is more challenging, but I think on NuScene, we are penalised less? Or?

Best, Shengyu

xinge008 commented 3 years ago

In fact, I do not try to fuse multiple sweeps in NuScene dataset. All these results are obtained via single key frame.

Generally, fusing multiple sweeps may benefit these tailed classes in the instance level, and copy-paste technique maybe a good alternative.

ShengyuH commented 3 years ago

okay great, thanks.