Photometric Invariant Feature Detection based on Oriented Tensor Filter

Traditional visual low-level features detection usually ignores color and photometric nature which can be utilize to exploit useful iso-luminance information and eliminate the unexpected shade-shading-specular effect. In this paper, we proposed a new feature detection method which integrated photometric quasi-invariant model with a new version of color tensor formed by a nonlinear filter named oriented tensor filter. We also investigated the relation between oriented tensor filter with popular tensor voting methodology in theory. Experiments show that photometric invariant features detected by our method, such as edge and corner are more effective and robust than tradition methods.