Multi-scale phase-based local features

摘要: Local feature methods suitable for image based object recognition and the estimation of motion structure are composed two steps, namely 'where' 'what' steps. The step (e.g., interest point detector) must select points that robustly localizable under common deformations whose neighborhoods relatively informative. local extractor) then provides a representation neighborhood is semi-invariant to deformations, but distinctive enough provide model identification. We present quantitative evaluation both steps three recent methods: a) phase-based features (Carneiro Jepson, 2002), b) differential invariants (Schmid Mohr, 1997), c) scale invariant transform (SIFT) (Lowe, 1999). Moreover, in order make approach more comparable other approaches, we also introduce new form multi-scale detector be used its step. results show lead better performance than approaches when dealing with illumination changes, 2D rotation, sub-pixel translation. On hand, somewhat sensitive large shear changes methods. Finally, demonstrate viability simple system.