How to voxelize point clouds like SpConv?

mit-han-lab / torchsparse

[MICRO'23, MLSys'22] TorchSparse: Efficient Training and Inference Framework for Sparse Convolution on GPUs.

https://torchsparse.mit.edu

MIT License

1.16k stars 132 forks source link

How to voxelize point clouds like SpConv? #233

Open WWW2323 opened 10 months ago

WWW2323 commented 10 months ago

Hi, thanks for your excellent work! I find when using torchsparse.utils.quantize for voxelization, each voxel only selects one point. In SpConv, each voxel can contain several points, which can reduce information loss of quantization. Is there any interface in torchsparse that can implement the similar function? Looking for your reply, thanks~