Toronto-3D and OpenGF dataset code for RandLA-Net

Code for Toronto-3D has been uploaded. Try it for building your own network.

Will release code for OpenGF later

Train and test RandLA-Net on Toronto-3D

1. Set up environment and compile the operations - exactly the same as the RandLA-Net environment
2. Create a folder called data and move the .ply files into data/Tronto_3D/original_ply/
3. Change parameters according to your preference in data_prepare_toronto3d.py and run to preprocess point clouds
4. Change parameters according to your preference in helper_tool.ply to build the network
5. Train the network by running python main_Toronto3D.py --mode train
6. Test and evaluate on L002 by running python main_Toronto3D.py --mode test --test_eval True
7. Modify the code to find a good parameter set or test on your own data

Sample results of Toronto-3D

The highest results reported are from RandLA-Net in Hu et al. (2021). Here are some results I got on my code with the default parameters. The largest factor in mIoU is the accuracy of Road Markings, which is impossible to be classified with XYZ only.