nvlabs/splatnet

SPLATNet is a research tool from NVIDIA that helps computers understand 3D point clouds, the raw 3D scan data produced by LiDAR sensors, depth cameras, and similar devices. Each point in a cloud represents a spot in physical space, and this software can automatically label what those points represent, such as identifying which parts of a scanned building are windows, walls, or doors. At a high level, the system processes point cloud data through a neural network built on a "sparse lattice" structure. That structure lets it efficiently organize and reason about scattered 3D points without needing them arranged in a regular grid. The project comes with pre-trained models and example workflows for two tasks: segmenting building facades into parts, and segmenting 3D object models (from a dataset called ShapeNet) into components like legs, seats, or backs. Someone working on autonomous vehicles, robotics, architecture, or 3D scanning might use this as a starting point. For example, if you have LiDAR scans of streets and want to automatically identify which points are road, sidewalk, or building, this type of approach is directly relevant. Similarly, if you are analyzing 3D scans of manufactured parts and need to separate them into sub-components automatically, the ShapeNet example shows how that works. The project is built on an older deep learning framework called Caffe and relies on a companion library called bilateralNN. It was published as a CVPR 2018 paper, and the license (CC BY-NC-SA 4.0) restricts it to non-commercial use. The README provides a Docker-based setup path, though it notes the included image was built for older GPUs and may need adjustment for newer hardware. This is a research codebase rather than a polished product, so expect some assembly required.