Patchlets: Representing Stereo Vision Data with Surface Elements

This paper describes a class of augmented surface elements which we call patchlets. Patchlets are planar surface elements generated from dense stereo vision 3D range images. Patchlets have a position, surface normal and size. In addition they have confidence measures on the position and normal direction that are based on the sensor accuracy. These confidence measures facilitate their use with probabilistic methods such as clustering for range image segmentation. Patchlets are formed by the projection of a pixel within the stereo image onto a sensed surface. They are surface elements that are constructed directly from the sensor data and can be used as a fundamental sensed-data primitive. We describe patchlet formation from the stereo disparity image, the propagation of errors from the stereo sensor model, and confirm experimentally the patchlet model representation. We provide surface segmentation as a sample patchlet application.