Helper functions and illustrations of COLA: COarse LAbel pre-training for 3D semantic segmentation of sparse LiDAR datasets.
Each dataset folder present the following configuration:
- DatasetName
- map_DatasetName_to_CL.py (can be launched to read (and optionnaly save) a frame for the target dataset with the coarse labels)
- table_DatasetName_to_CL.py (map the original label of the target dataset to the coarse labels)
- relabel_DatasetName_to_CL.py (relabel a full sequence of a dataset)
An example to save Coarse Labels from the frame 10 of the sequence 03 SemanticKITTI:
python SemanticKITTI/map_SemanticKITTI_to_CL.py -f 10 -s 3 -d semanticKITTIDirectory --saved 1
An other example to process the full sequence 03 of SemanticKITTI:
python SemanticKITTI/relabel_SemanticKITTI_to_CL.py -s 3 -d semanticKITTIDirectory
Link for the SRU-Net and SPVCNN pretrained model (with SemanticPOSS as the target): https://cloud.mines-paristech.fr/index.php/s/555t3DzzMmhA2xp
SRU-Net is the default model from : https://github.com/facebookresearch/PointContrast
SPVCNN is the default model from : https://github.com/mit-han-lab/spvnas