search

ma-xu / pointMLP-pytorch

[ICLR 2022 poster] Official PyTorch implementation of "Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework"
Apache License 2.0
485 stars 61 forks source link

Questions about training materials #28

Closed aaaa47080 closed 2 years ago

aaaa47080 commented 2 years ago

　　Whether it is ModelNet40 or objectdateset, their xyz starting coordinates seem to have been adjusted, such as taking the center of the object as (0,0,0) and normalizing, but pointnet++ is based on Farthest Point Sampling for sampling, so in fact our The coordinate axis, whether it is set to any point in space as the starting point (0, 0, 0) relative to the change of x, y, z, will not affect the model, because in terms of classification, the model is based on the 3D object The outer contour of the point cloud to determine what the object is (MLP based invariance)? (Of course, no matter how the coordinate axes are converted, they must be normalized)

　　I would like to ask the boss, what do you think about this and whether it will affect PointMLP. Looking forward to your reply.

ma-xu commented 2 years ago

@aaaa47080 Sorry that I chould not understand your questions clearly.

Since PointMLP is similar to PointNet++ (both MLP-based, same data processing, normalization layers, etc.), they should share the same properties.

aaaa47080 commented 2 years ago

Sorry sir, maybe I didn't think clearly.

What I really want to ask is the feature data you feed into the model, is it the filtered and normalized point cloud x, y, z?

Since I am interested after seeing your literature, I would like to do some academic research.

However, the point cloud data set currently collected at hand consists of three features, which are ply files. Is it possible for me to feed the model directly.

But I don't know how to preprocess these data. So I want to ask your method, or advice.