search

IRMVLab / RegFormer

[ICCV2023]RegFormer: An Efficient Projection-Aware Transformer Network for Large-Scale Point Cloud Registration
100 stars 5 forks source link

cpu only #16

Open darkdelting opened 2 months ago

darkdelting commented 2 months ago

I want cpu only to test pretrained model, but pointnet and CUDA-based KNN searching c++lib need change. if i use https://github.com/yanx27/Pointnet_Pointnet2_pytorch/tree/master cpu only modle ， can it replace the current pointnet ?

liujiuming123 commented 2 months ago

Since our feature extraction part follows the same pattern like EfficientLO (TPAMI22), they utilize CUDA for point feature subtraction in parallel. Maybe using cpu is not a good choice I think, mostly for efficiency consideration.

darkdelting commented 2 months ago

Thank you for your answer!!! but cpu is cheap ~ ^ ^ ~