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
854 stars 180 forks source link

Thanks for your great work. But it seems that there are many mismatches in the code. #136

Closed l9761116 closed 2 years ago

l9761116 commented 2 years ago

I'm new to AI algos. So it's a bit difficult for me when the code cannot run successfully, to find the problems and modify them. Since your work is such a hot choice in Semantic KITTI task, I really see the good if these issues can be fixed.

Basically, I met three biggest problems, all in 'segmentator_3d_asymm_spconv.py'.

First, 

x.feature = Func(x.feature)

is an incompatible line. There are many of these lines in the file.

Second, I met an error

ValueError: subm with same indice_key must have same kernel size, expect [1, 3, 3], this layer [3, 1, 3]

I looked through the issues and found that someone else also met this problem. I guess this is because you have uploaded different versions of the algorithm. But I'd appreciate if you could fix this mismatch.

Last, since spconv 1.x is deprecated now, it'll be great to update the spconv methods names.

Generally, this work is very helpful to me. Thanks for your great work!

xinge008 commented 2 years ago

I do not use the spconv >= 2.x and there is no pretrained model for spconv>=2.x; In my opinion, in spconv>=2.0, some basic definitions are changed so you need fit it with some modifications.