search

jac99 / MinkLoc3D

MinkLoc3D: Point Cloud Based Large-Scale Place Recognition
MIT License
122 stars 19 forks source link

Complexity Analysis #9

Closed BIT-MJY closed 1 year ago

BIT-MJY commented 1 year ago

Hi @jac99 ,

Thanks for your great open-source code! Currently I want to compare our method with MinkLoc3D on FLOPs, but I find it difficult to calculate FLOPs for the MinkowskiEngine model. Have you ever achieved this using some existing tools or your ad-hoc function?

jac99 commented 1 year ago

Hi, unfortunately, I haven't done any FLOPS analysis of MinkLoc3D method. I've seen a similar question posted in MinkowskiEngine repo: https://github.com/NVIDIA/MinkowskiEngine/issues/115, but there was no answer. MinkowskiEngine uses some clever hashing method to quickly find values of neighborhood voxels in order to efficiently compute sparse convolutions. So in the best case retrieving voxel values is constant, but in the worst case it is linear with respect to the number of points. Maybe take a more pragmatic approach, benchmark the methods on the same hardware, and report processing time per point cloud.

BIT-MJY commented 1 year ago

Get it. Thanks for your suggestion.