Network Training

Saved Models

• Runs

  • This directory contains TensorBoard log files for the training and validation of each model.
  • Models were chosen by performance on the mean intersection over union (mIoU) metric of the validation set.
  • Training was generally 24 to 48 hours, but stopped early if the model began to overfit the data.

• Weights

  • Contains the best weights for each model, determined by performance on validation set mIoU.

Training

• train_carla.py

  • Script for training models on CarlaSC dataset.
  • Parameters are set at the top, including which network to train, where the data is located, and whether to remap the classes from 23 to 11.
  • Logs training data, and saves frozen weights each epoch.

• train_kitti.py

  • Script for training models on KITTI dataset.
  • Operates similarly to training file cor CarlaSC.

• dataset.py

  • Data loader for the CarlaSC data set.
  • Loads data into raw point clouds or voxel grids, and applies data transformations if any. Also remaps semantic labels if required.

• utils.py

  • General utilities for training or testing.
  • Contains functions to visualize predictions, load models or resample free space for SSCNet training.

Testing

• test_carla.py

  • Script for testing models on CarlaSC dataset.
  • Parameters are set at the top, including which network to train, where the data is located, and whether to remap the classes from 23 to 11.
  • Boolean flags are used to set which test to run.

• test_kitti.py

  • Script for testing models on KITTI dataset.
  • Saves results in SemanticKITTI format for submission.