search

QingyongHu / RandLA-Net

🔥RandLA-Net in Tensorflow (CVPR 2020, Oral & IEEE TPAMI 2021)
Other
1.3k stars 320 forks source link

Questions about the voting #61

Open linhaojia13 opened 4 years ago

linhaojia13 commented 4 years ago

Hi, I notice that you use the voting in the evaluation. It seems that voting is a common trick in 3d semantic segmentation, which can be found in many other method, such as PointConv, PointASNL, FPConv. However, I am having trouble in undertanding the voting in semantic segmentation task. For classification, we can rotate the object for different angles and get the final result by comprehensively consider the scores output in different angles. The rotation is the very vote in the classification. But when I look into the evaluation codes for semseg, it seems that it just conduct 'num_votes' times same inference. Could you please tell me what is the vote for semseg?

nlgranger commented 3 years ago

Random guess as I just read through the paper and stumbled on this issue: isn't it to average over different point sub-samplings?