A point cloud is an unstructured set of 3D points — each defined by (x, y, z) coordinates and optionally colour, intensity, or normals — captured by LiDAR sensors, depth cameras, or photogrammetry, serving as a raw spatial representation of a physical scene or object.
Point clouds are the primary output of 3D scanning technologies. A terrestrial LiDAR scan of a building can contain billions of points; a depth camera frame yields tens of thousands. Unlike meshes, point clouds have no connectivity information — they are raw, unordered samples of surfaces.
Processing point clouds with deep learning requires specialised architectures. PointNet and its successors operate directly on unordered point sets; sparse 3D convolutions (MinkowskiNet, TorchSparse) voxelise the space and apply efficient convolutions; and transformer-based methods (Point Transformer) use self-attention over local neighbourhoods.
Common tasks include semantic segmentation (labelling each point), classification, registration (aligning scans), and surface reconstruction (converting points into a mesh or splat representation). Clean-up — removing noise, outliers, and scanner artefacts — is a critical preprocessing step.
Datameister builds 3D deep learning pipelines that process point clouds from LiDAR and depth sensors for industrial inspection, construction monitoring, and robotic scene understanding.
Datameister's presence at ECCV 2024 in Milan was a deep dive into the latest advancements in Computer Vision. This blogpost explores groundbreaking techniques like 3D Gaussian Splatting and its enhancements, such as WildGaussians and Gaussian Frosting, showcasing their potential to revolutionize real-time rendering and scene representation.
Read Article →