Question about the difference between PVCNN and SPVCNN

[jialeli1]

Awesome performance!

If the NAS part is excluded, is it just that the sparse 3D convolution is added to SPVCNN, compared to PVCNN? However, I thought you already used the sparse 3D convolution in PVCNN when I read PVCNN some time ago. After all, you have already cited the sparse 3D convolution and SECOND in PVCNN.

If the answer is yes, I feel very regret that I missed this idea.

[kentang-mit]

Hello,

We actually didn't use sparse 3D convolution in PVCNN. Besides this apparent difference, another important difference in SPVCNN and PVCNN is that the voxelization / devoxelization operations are more challenging, therefore we don't want to have too many SPVConv layers.

Best,
Haotian

[chaitjo]

Hi @kentangSJTU, I had a quick follow-up on this discussion:

the voxelization / devoxelization operations are more challenging, therefore we don't want to have too many SPVConv layers

Do you mean that the voxelization/devoxelization operations for sparse tensors are more time costly in the SPVConv layers compared to PVConv (i.e. dense 3D convolution)? Any further insight would be very helpful.

[kentang-mit]

Hi @chaitjo,

Thanks for your interest. For dense tensors, we can directly infer a point's memory location given its coordinates (coordinate x,y,z will correspond to memory location x * y_max * z_max + y * z_max + z). However, for sparse tensors, there is no such property, and we have to rely on hash map query to obtain the memory location of a point given its coordinates. Therefore, voxelization and devoxelization operations can be slower.

Best,
Haotian

[chaitjo]

Thanks for the explanation!