maxsampson/kpconv-pdal

Derivative work of https://github.com/HuguesTHOMAS/KPConv-PyTorch and https://github.com/chambbj/kpconv-pdal, aimed first at evaluating US3D dataset, and now being used as an introduction with the Autzen dataset. The initial classes to be evaluated on the Autzen dataset are ground, vegetation, and buildings.

We'll add the requisite LICENSE file in subsequent commits, but as with the source repository (https://github.com/HuguesTHOMAS/KPConv-PyTorch/blame/master/README.md#L55) this work will be released under the MIT license.

To create a Conda environment to run kpconv-pdal, run the following command (assuming you have Conda installed of course) from within the root directory of the repository.

conda env create -f environment.yml
conda activate kpconv-pdal-env

We currently still use some of the C++ extensions provided in the original KPConv-Pytorch release. To compile these, from within the cpp_wrappers directory, issue the following command.

sh compile_wrappers.sh

At this point, you should be able to train models and predict classification labels. After modifying the LASConfig class within train_LAS.py, namely to adjust paths to the dataset and the tensorboard run directory, simply execute the following.

python train_LAS.py

The current means of predicting labels is to modify test_LAS.py to point to the results log directory, and to uncomment line 42 of datasets/LAS.py. You can then run the following.

python test_LAS.py