RandLA-Net

Why we included this project

RandLA-Net is the official TensorFlow implementation of a CVPR 2020 paper on semantic segmentation of large-scale 3D point clouds, so it assigns a class label to every point in dense LiDAR or scan data rather than just finding objects in an image. Its core trick is random sampling paired with a local feature aggregation module, which keeps memory and compute bounded enough to handle million-point clouds that simpler point-wise networks choke on. If you are labeling autonomous driving LiDAR or working with datasets like SemanticKITTI, Semantic3D, or S3DIS, you get training and evaluation scripts plus pre-trained weights to reproduce the published results. Expect some setup work: it targets TensorFlow 1.x and a 2020-era environment, so this is a research codebase to adapt, not something to drop into a production pipeline. It still works well as a reference baseline when you are prototyping point cloud segmentation and want a result you can compare against the literature.