Octree Mapping from a Depth Camera

To render artificial objects with consistent shading from arbitrary perspectives, a 3D scene needs to be constructed from the camera frames. Data parallel GPU computing allows for real-time 3D mapping of scenes from depth cameras such as the Kinect sensor. Noise in the camera’s depth measurements can be filtered over multiple image frames by representing the scene as a voxel-based map rather than as a collection of raw point clouds. However, a dense voxel grid representation is not suitable for large scenes or live rendering for use in games.