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. |
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.
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 :)