This repository propose python scripts for Semantic Segmentation with Transfer Learning for MLS Point Clouds. The library is based on the project ConvPoint
Semantic data can be downloaded at https://semantic3d.net.
In the folder semantic3D_utils:
python setup.py install --home="."
Then, run the generation script:
python tummls_prepare_data.py --rootdir /media/liangdao/DATA/Paris_and_Lille --savedir /media/liangdao/DATA/Paris_and_Lille
python semantic3d_prepare_data.py --rootdir /media/liangdao/DATA/small/area123 --savedir /media/liangdao/DATA/small/convpoint
python tummls_prepare_data.py --rootdir /media/liangdao/DATA/small/subarea --savedir /media/liangdao/DATA/small/subarea
TUM-MLS can be downloaded at testdaten
The training script is called using:
python semantic3d_seg.py --rootdir your_pointcloud_path --savedir your_save_folder_path
e.g.
python semantic3d_seg.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data
Add --continuetrain at the end, which means reading a pretrained model from savedir and continue update the parameters
python semantic3d_seg.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/SegBig_8192_nocolorTrue_drop0.5_2022-08-20-17-52-27 --continuetrain
python semantic3d_seg.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/SegSmall_8192_nocolorTrue_drop0.5_2022-06-06-22-30-46 --continuetrain
Sematnic3D training, area1 test
python semantic3d_seg.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SegBig_8192_nocolorTrue_drop0.5_2022-09-14-09-23-17 --test --savepts
python semantic3d_seg_trans.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/
python semantic3d_seg_gan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor --finetuning
python semantic3d_seg_gan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_Domain_8192_nocolorTrue_drop0.5_2022-08-01-17-55-41 --continuetrain
python semantic3d_seg_gan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_Domain_8192_nocolorTrue_drop0.5_2022-08-01-17-55-41/SegBig_Domain_8192_nocolorTrue_drop0.5_2022-08-02-16-12-18 --test --savepts
python semantic3d_seg_finetuning.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_8192_finetuning_linearlayer_nocolorTrue_drop0.5_2022-08-30-17-17-00 --finetuning
python semantic3d_seg_finetuning.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor --finetuning
python semantic3d_seg_childtuning.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/ --finetuning
python semantic3d_seg_childtuning.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/ --finetuning
python semantic3d_seg_mmd.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor --finetuning
update discriminator when semantic3D training; load pretrained model python semantic3d_seg_gan_reverse.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/ --finetuning
python semantic3d_seg_gan_unsup_step1.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-07-23-21-23-51 --continuetrain
python semantic3d_seg_gan_unsup.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor --finetuning
python semantic3d_seg_gan_unsup.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-09-30-20-00-25 --continuetrain
python semantic3d_seg_gan_unsup.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-07-23-21-23-51/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-08-25-18-44-10 --test --savepts
'''
python semantic3d_seg_gan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor --finetuning
python semantic3d_seg_gan_pointdan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-07-23-21-23-51 --continuetrain
python semantic3d_seg_gan_pointdan.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_GAN_8192_nocolorTrue_drop0.5_2022-09-25-18-33-05 --test --savepts
To predict on the test set (voxelized pointcloud):
python semantic3d_seg.py --rootdir /media/liangdao/DATA/segmentation/ConvPoint/data/Prepare/train/pointcloud --savedir /media/liangdao/DATA/segmentation/ConvPoint/data/SEMANTIC3D/SegBig_nocolor/SegBig_8192_childtuning_nocolorTrue_drop0.5_2022-08-30-17-20-24 --test --savepts
python semantic3d_seg.py --rootdir /media/liangdao/DATA/origin_data/convpoint/test/pointcloud/ --savedir /media/liangdao/DATA/origin_data/convpoint/SegBig_8192_nocolorNone_drop0.5_2022-04-28-02-36-49 --test
Finally to generate the prediction files at benchmark format (may take som time):
python semantic3d_benchmark_gen.py --testdir path_to_original_test_data --savedir /path_to_save_dir_benchmark --refdata path_to_data_processed --reflabel path_to_prediction_dir
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.
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.