About CUDA kernel error and some questions.
Opened this issue · 4 comments
Deleted user commented
Hello, thank you for providing the code.
I just started learning about pointcloud and there are a few things I don't understand.
(1). What is the role of embedding layer in PointMLP?
(2). Which part is the Geometric Affine Module in code? Also, in Figure 6 of the paper, how many times the Geometric Affine Module appeared?
(3). Do you learn all features in one process with the entire point cloud as input, or n selected point features?
That's all I ask. Thank you very much.
ma-xu commented
Hi @popopochan
Thanks for your interests, here are the responses.
- just mapping 3-d xyz to high dimensional features.
- See the LocalGrouper module in our codes. 4 times, each stage has one.
- Sry that I didn't get it. We feed all the 1024 points into the network for feature extraction (classification task).
Feel free to let mw know if you have any further questions.
Deleted user commented
Hi, @ma-xu
Thank you for your timely reply.
I still have some questions.
- I understand that this module mapping in a low dimension to high dimension, but what is the reason (purpose) for embedding?
- I understand. Is this the affine module before the preblock at all stage?
- It's a neighborhood-by-neighbour process, but at the end it’s extracting all the features collectively?
Thank you.
ma-xu commented
- Directly operating on 3-dimensional features may not achieve good performance.
- Yes.
- If I correctly understand your question (operating locally in each stage), the answer is yes.
Deleted user commented
OK, thank you.