1. https://github.com/sj-li/minet

    1. flops-counter.pytorch (https://github.com/sovrasov/flops-counter.pytorch)
    2. flops of k-nn, knn

  2. Ranked #2 on Real-Time 3D Semantic Segmentation on SemanticKITTI

    Untitled

  3. Ranked #19 on 3D Semantic Segmentation on SemanticKITTI: 55.2%

    1. #1 is 2DPass, 72.9%

    Untitled

Network

Untitled

  1. 3-scale: 最大的用MobileBlok，最小的用BasicBlock；Table V探索了如何取舍以取得最大性价比。
  2. TABLE II说明，UFM最重要，interaction稍微弱一点，MFM最弱。
  3. Losses：除了在中间合适的尺度上加，loss的type也值得关注
    1. supervision to intermediate parts: weighted cross entropy loss Ls + edge supervision loss Le;
    2. end of network: weighted cross entropy loss: Lls + IOU loss: Lfs;

References

  1. Multi-scale Interaction for Real-time LiDAR Data Segmentation on an Embedded Platform, IRAL21