Overarching repository for LiDAR Segmentation with HD [In progress]
Each baseline has it's own docker image, to run simply start instance with the respective image
Can be run with the package cenet_image:1.0 in this repository. To start compiling run the following command:
python train_tls.py -d /root/dataset/tls/dense_dataset_semantic/ -ac config/arch/tls.yml -n res
Can be run with the package cenet_image:1.3 in this repository. To start compiling run the following command:
sh tf_interpolate_compile.sh # at /SQN/tf_ops/3d_interpolation
python data_prepare_personalized.py # at /SQN/utils
python main_TLS_kitti.py --mode train --gpu 0 --labeled_point 100%
Can be run with the package cylinder3d_image:1.3 in this repository. To start compiling run the following command:
python -u train_cylinder_asym_tls.py 2>&1 | tee cylinder_logs_tee.txt #training
python demo_folder.py --demo-folder /root/dataset/dense_dataset_semantic/sequences/01/velodyne/ --save-folder ./save --demo-label-folder /root/dataset/dense_dataset_semantic/sequences/01/labels/ # testing?
Make sure to move trained model from model_save_dir to model_load_dir once the training is done. And config/TLS.yaml needs to be setup correctly.