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
858 stars 180 forks source link

Why result in CodaLab competitions is different from result in the paper？ #139

Open 91916117qq opened 2 years ago

91916117qq commented 2 years ago

Dear author: I am a novice in the field of 3d semantic segmentation. I notice your miou in CodaLab competitions is 68.9, but in your paper the result is 67.8. Why are the two results different? Thanks.

L-Reichardt commented 2 years ago

The SemanticKITTI benchmark lists 67.8

As a sidenode, this repository covers the implementation of C3D without the 2nd stage of the pointwise refinement module, or the weighted Cross entropy loss. The applicable mIOU to reference to is 64.3 from the authors other paper