train

LiDAR-Bonnetal Training

This part of the framework deals with the training of segmentation networks for point cloud data using range images.

Tasks

• Semantic Segmentation.
• Panoptic Segmentation [Soon].

Dependencies

First you need to install the nvidia driver and CUDA, so have fun!

• CUDA Installation guide: link

• System dependencies:

  $ sudo apt-get update 
  $ sudo apt-get install -yqq  build-essential ninja-build \
    python3-dev python3-pip apt-utils curl git cmake unzip autoconf autogen \
    libtool mlocate zlib1g-dev python3-numpy python3-wheel wget \
    software-properties-common openjdk-8-jdk libpng-dev  \
    libxft-dev ffmpeg python3-pyqt5.qtopengl
  $ sudo updatedb

• Python dependencies

  • first apt install python3-tk python3-distutils

  • $ sudo pip3 install -r requirements.txt

  • install torch and torchvision. In order case they must be special versions, so must be installed manually (see requirements.txt file)

Running

• active the venv in which requirements were installed

• cd train/tasks/semantic/

• For inference:

  python infer.py -d __pointcloud_folder__ -m __model_folder__ -ac __arch_cfg_yaml__

  • __pointcloud_folder__ must be a path to a folder containing pointclouds recorded as .bin or .pcd. its structure must be __pointcloud_folder__/sequences/xx/velodyne/... where 'xx' are numbers of sequences to be used (and corresponds to the those configured in data_cfg)
  • __model_folder__should contain the model and config files
  • __arch_cfg_yaml__ should be the name of the arch config file to be used, iof not using the default 'arch_cfg.yaml'
  • see python infer.py -h for more args

  In the end of the of the inference, the script will print the path to the log folder where the inference results were saved

• Visualize inference results:

  python visualize.py -d __pointcloud_folder__ -s 00 -ac __arch_cfg_yaml__ -p __log_folder__

  • __pointcloud_folder__ and ``arch_cfg_yaml` same as for inference
  • __log_folder__ is the folder with the inference results