3D grouping by viewpoint consistency ascent

摘要

The viewpoint consistency constraint (VCC) provides a powerful way to discover extended feature groups, and to test hypotheses in object recognition. Lowe's incremental method fails in complex scenes, and an exhaustive tree search (e.g. Grimson and Lozano-Perez) is too expensive. We present a state space approach in which transitions are made which monotonically ascend a measure of viewpoint consistency.