Torch Points3D

Full post: Toward data science

With the rise of ever more affordable LiDAR sensors and more efficient photogrammetry algorithms, 3D point cloud data have become easier than ever to acquire. The deep learning community has embraced this trend by developing new network architectures to perform various tasks on 3D data. The sheer volume of possibilities (data layouts, data augmentation, convolution strategies, etc…) is such that it can be time-consuming to find the best fit for your data or problem.

Our framework, Torch Points3D, was developed to become the torchvision of point cloud data: a flexible and extensible framework for researchers and engineers alike working on point cloud-based machine vision. Ever wondered how KPConv could perform for point cloud registration? Or PointNet++ for object detection with random sampling instead of furthest point sampling as suggest in RandLa-Net? With Torch Points3D you can now try multiple state-of-the-art backbone models in just a few lines of code.

Ever wondered how KPConv could perform for point cloud registration? Or PointNet++ for object detection with random sampling instead of furthest point sampling as suggest in RandLa-Net? With Torch Points3D you can now try multiple state-of-the-art backbone models in just a few lines of code. After a quick recap about the specificity of point clouds, we will present the following aspects of Torch Points3D:

  1. Optimized data layouts for point cloud data
  2. Native integration of many academic datasets
  3. Fast and robust data processing and data augmentation
  4. Tested convolution kernels for a range of sparse and point-based architectures
  5. Easy-to-use API for accessing datasets, data transforms, and preconfigured models

We provide training scripts with useful features such as model checkpointing, logging to Tensorboard and Weight and Biases, easy configuration of hyper parameters with Facebook's Hydra, and more. You can also use our core components with your favourite training framework such as PyTorchLightning, for example. We hope you enjoy the read! We are open to your feedback and contributions here.

We hope you enjoy the read! We are open to your feedback and contributions here.

System diagram of Torch Points3D, data flow highlighted in red

Principia Labs Ltd is registered in the United Kingdom. Company number 12371387.