Master's Thesis about 3D LiDAR Semantic Segmentation using Range Images and Retentive Networks.

The purpose of the Thesis is the development of a lightweight model that is able to achieve real-time performance while requiring low amount of memory and an optimized training process.

• Pytorch
• yaml
• tqdm

Download the files from the SemanticKITTI website

./semanticKITTI/
├── 
├── ...
└── dataset/
    ├──sequences
        ├── 00/           
        │   ├── velodyne/	
        |   |	├── 000000.bin
        |   |	├── 000001.bin
        |   |	└── ...
        │   └── labels/ 
        |       ├── 000000.label
        |       ├── 000001.label
        |       └── ...
        ├── 08/ # for validation
        ├── 11/ # 11-21 for testing
        └── 21/
	    └── ...

Run the following commands, specifying the SemanticKITTI dataset path

cd RangeRet
python3 main.py <config_path> <data_path>

Example: python3 main.py config/RangeRet-semantickitti.yaml /semanticKITTI/dataset/

Checkpoint will be saved in checkpoints/model-checkpoint.pt, the final model will be saved as rangeret-model.pt and training logs in log/ with the current date and time.

Run the following script, specifying the SemanticKITTI dataset path

cd RangeRet
python3 infer.py <config_path> <model_path> <data_path> <pred_path> <split>

where split can be train, valid or test.

Example of inference on SemanticKITTI validation set: python3 infer.py config/RangeRet-semantickitti.yaml rangeret-model.pt /semanticKITTI/dataset predictions/ valid

To correctly evaluate the results, please refer to the scripts available at semantic-kitti-api.

Models trained on a subest of SemanticKITTI with 5000 scans, equals to the 25% of the entire dataset, and evaluated on the complete validation set, sequence 08.

Master Thesis document is available here

Code is built on RangeNet++ and Torchscale.