RIU-Net for point cloud semantic segmentaion

• This work represents the implementation of the RIU-Net: Embarrassingly simple semantic segmentation of 3D LiDAR point cloud paper with the SemanticKITTI dataset related to semantic segmentation of point clouds made for autonomous cars.
• Paper data for this implementation (can change depending on the context of work):
  • Loss function: CrossEntropyLoss
  • Batch size: 8
  • Epochs: 10
  • Learning rate: 0.001
  • Batch normalization with momentum = 0.99
  • Optimizer: Adam or SGD
• Current status: Getting there

Important considerations:

• Using pytorch functions.
• The network trains with 2D images of point clouds, which means that some pre-processing of the dataset was done, such as:
  • Spherical projection of the .bin and .label archives.
  • Split the pojections into two kinds of image values: Reflectance and depth (2 channels image), and label index (1 channel image).
  • The label dictionary can be found at the semantic kitti api repository and the learning_map dictionary is the one used in this work.
  • Image dimensions set to 64x1024 to match the dataset.
  • Since this work is done using Google Colab notebooks, the dataset was uploaded on google drive beforehand.
• All the pre-processing and visualization was done using the Cloud2DImageConverter repository
• Any other information about this work can be found on the notebooks text cells as well as the visualization tutorial.