range_net

How to add a point cloud

You need to have a point cloud with the three coordinates x, y, z and the verticality.

1. Open it in cloud compare and cut the point cloud to save only the admissible centers (the ones that will be selected afterwards). To do so, use to tool called segment.
2. Save the admissible centers as binary in data/centers.
3. Go to the city_statistics notebook, you will need to change to absolute path so that the jupyter kernel is in the top folder of the repository.
4. Change the point_cloud_name and run all cells. You can then add the new values obtained from the output's cells to the init file. In total, the dictionaries CENTERS, Z_GROUNDS and ROTATIONS need to be updated.
5. You can check the visualizations cells of the notebook to see if everything is happening fine.
6. To use RangeNet++ afterwards, you need to update the dictionnary CITY_INFERANCE_FOLDER in the init file.

You can now freely generate new samples from the point cloud that you have. For that, you need to use the command line create_dataset. If by any chance the pipeline gets broken, a notebook called create_dataset has been made to help debugging.

How to train RangeNet++

You can open the Jupyter Notebook directly in colab by clicking here: Consider restarting the runtime if a module is not found.

The step are explained in the notebook. You need to upload your data into the Colab session. The way it is done here is by using Google Drive.

For further details, please go and check the documentation of RangeNet++.

How to infer and merge the predictions

The steps to follow are the following:

1. First generate the samples you would like to predict on by using the command line create_dataset.
2. Then generate the predictions with the weights that you have trained from the previous section.
3. Merge the predictions by using the command line merge_labels.

It has been coded here: Consider restarting the runtime if a module is not found.