theovincent / 3dpointcloudclassification

Challenge to classify 3D point clouds of cities into Ground - Building - Poles - Pedestrians - Cars - Vegetation

3DPointCloudClassification

Challenge to classify 3D point clouds of cities into Ground - Building - Poles - Pedestrians - Cars - Vegetation. The challenge description and the dataset are available here. The best results ended can be summarized as follows, with IoU for Intersection over Union:

	Average	Cars	Pedestrians	Ground	Building	Vegetation	Pole
Rank	3	1	3	1	3	3	3
IoU	53.8	85.5	5.5	97.7	77.1	38.1	19.0

The approach considered here is coming from the RangeNet++ paper. This time the algorithm is applied on outdoor point cloud instead of LiDAR scans. I advise the reader to have a look to the report, that I have made, that explains how this transfer is done. You can see here the overall pipeline:

The outdoor pipeline is first split into virtual LiDAR scans (far left). Then each virtual scan is passed throught the RangeNet++ separately (middle left to middle right). Finally, the predictions are merged together into the original point cloud (far right). Here 1000 virtual scans have been used to get the predictions.

Animation showing on the test point cloud how the virtual scans gradually help to label the entire dataset.

Organization

This repository is organised into two main folders:

• classifier_3D is the folder where feature classification can be done. Among them, you can use: verticality, linearity, planarity, sphericity, x, y, z.
• range_net is the folder where the transfer from outdoor point cloud to LiDAR is being done. Feel free to have a look to the README.md file of that folder to have a more complete description.

How to install

You can install this repository in your personnal computer by cloning it and installing the package in editable mode:

pip install -e .

If you want to use a Docker container with Visual Studio Code, a .devcontainer has been created for you :)

References

RangeNet++

A. Milioto, I. Vizzo, J. Behley, and C. Stachniss. RangeNet++: Fast and Accurate LiDAR Semantic Segmentation. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019.