How to train 3D point cloud data? Where should I beginn?

Question

I'm an automotive engineer student, and at the moment I'm working in a project for an autonomous bus at the university with 3D point clouds from a lidar sensor. My job here is to train the point cloud with deep learning algorithms. But I do not know exactly how to start. I found many sources on the internet. But it is also too diverse for me as a beginner, I do not know how to start first. Can someone give me some tips? Or good source for beginners.

Thank you in advance!

AFAIK, deep learning on point clouds is not settled yet on a 'canonical' way to go (Sparse convolutions vs PointNet-derived vs other less common for now). You can search for most recent lectures or presentations with overview of this topic and then follow links. Look for example https://www.qwertee.io/blog/deep-learning-with-point-clouds/. — datjko, Oct 27 '19 at 17:43

score 1 · Answer 1 · answered Nov 19 '20 at 14:11

According to your question, I assume that you are working on your own LiDAR point cloud data rather than the public datasets. Initially, I would suggest annotating the data with 3D bounding boxes on point cloud data.

As concerned about deep learning algorithms I would prefer understanding Complex YOLO, VoxelNet, and PointNet. To understand the implementation PointNet algorithm you can refer to the Keras webpage trained on the ModelNet10 dataset.

score 0 · Answer 2 · answered Feb 24 '20 at 07:34

There are two possible approaches:

Feature engineering (user-defined global and/or local features)
Feature learning (models such as PointNet, PointNet++, and VoxNet)

If the feature engineering approach is selected, you can train your model with several different classifiers, such as Multilayer Perceptron (MLP), Support Vector Machines (SVM), and Random Forest (RF), for instance.

How to train 3D point cloud data? Where should I beginn?

2 Answers2