jbehley
commented
2 months ago The kNN majority vote helps only if the majority of points nearby have the correct class (which is our assumption); therefore, the "blue dots" don't have a majority of "correct" labels in the surroundings.
The "label bleeding" is a particular problem of the projective approaches and therefore using something like a point-based method or voxel-based approach might help to resolve such issues. There are also more modern approaches, like Cylinder3D or MinkowskiNet that might be more suited for your problem.
However, I cannot provide you any more specific help as it seems not related to RangeNet++.
Thank you for the great work. i want know how to use KNN,I set KNN-use to True in arch_cfg.yaml and infer I've tested several sets of KNN parameters, but I don't even see a change after using KNN By the way, this is my rangeview, I want to remove the blue dots and change them to pink dots, can you give me some advice on how to do that.