Reflectance-based surface saliency

Abstract

In this paper, we propose an original methodology allowing the computation of the saliency maps for high dimensional RTI data (Reflectance Transformation Imaging). Unlike most of the classical methods, our approach aims at devising an intrinsic visual saliency of the surface, independent of the sensor (image) and the geometry of the scene (light-object-camera). From RTI data, we use the DMD (Discrete Modal Decomposition) technique for the angular reflectance reconstruction, which we extend by a new transformation on the modal basis enabling a rotation-invariant representation of reconstructed reflectances. This orientation-invariance of the resulting reflectance shapes fosters a robust estimation of saliency maps linked to the local visual appearance behaviour of surfaces on the scene. The proposed methodology has been tested and validated on real surfaces with controlled singularities, and the results demonstrated its efficiency since the estimated saliency maps show strong correlation with sensorial visual assessments.