npm3d

NPM3D example

Data

Data can be downloaded at http://npm3d.fr/paris-lille-3d.

The script npm3d_prepare_data.py can be used to split the point clouds and save them at numpy format.

For the training set:

python npm3d_prepare_data.py --rootdir path_to_data --destdir path_to_data_processed

For the test set:

python npm3d_prepare_data.py --rootdir path_to_data --destdir path_to_data_processed --test

Training

python npm3d_seg.py --rootdir path_to_data_dir --savedir path_to_save_dir

python npm3d_seg.py --rootdir path_to_data_dir --savedir path_to_save_dir --nocolor

Testing

python npm3d_seg.py --rootdir path_to_data_dir --savedir path_to_save_dir --test

If the model was trained without using the lidar intensity:

python npm3d_seg.py --rootdir path_to_data_dir --savedir path_to_save_dir --test --nocolor

note: the test_step parameter is set 0.8. It is possible to change it. A smaller step of sliding window would produce better segmentation at a the cost of a longer computation time.

Fusion model

Once models (RGB and without color information) have been trained, it is possible to train a fusion model.

Training

python npm33d_seg_fusion.py --rootdir path_to_data_processed --savedir path_to_save_dirctory --model_rgb path_to_rgb_model_directory --model_noc path_to_no_color_model_directory

Test

python npm3d_seg_fusion.py --rootdir path_to_data_processeed --savedir path_to_save_dirctory --model_rgb path_to_rgb_model_directory --model_noc path_to_no_color_model_directory --test --savepts

Pretrained models

Pretrained models can be found here.

Note: due to change of affiliation and loss of data, these models are given as they are, without any performance guarantee. Particularly, they may not be the ones used in the final version of the paper.