search

Pointcept / PointTransformerV2

[NeurIPS'22] An official PyTorch implementation of PTv2.
352 stars 24 forks source link

modelnet40 voxelization #29

Closed JJBUP closed 1 year ago

JJBUP commented 1 year ago

Why does the ModelNet40 need voxelization during training and not during validation and testing, while S3DIS training and validation and testing need voxelization. My question appears in config/ base/ modelnet40.py. We look forward to your reply Thank you!

Gofinge commented 1 year ago

Hi, because ModelNet40 has 10,000 points/data, while S3DIS has 1,000,000 points/scene.

JJBUP commented 1 year ago

Therefore, we use the voxelized data for training, and the non-voxelized data for validation and testing. Will different densities of data not affect the validation and testing results？🧐

Gofinge commented 1 year ago

Maybe the instruction here can solve you double: https://github.com/Gofinge/PointTransformerV2#testing;

Also, another special tech which enabled here can acquire precise eval results during training: https://github.com/Gofinge/PointTransformerV2/blob/main/configs/scannet/semseg-ptv2m2-2-precise-evaluate.py