Shading cues for object class detection

Abstract

Recognition of object classes in natural images has made tremendous progress in recent years. Today's approaches often rely on powerful learning approaches as well as robust local 2D shape or appearance features. Exploiting 3D shape cues however has become unfashionable in recent literature. While shading cues play a major role in human perception of object shape, shape-from-shading techniques are seldom used today for object class detection. Drawing on ideas from the early days in object recognition this paper aims to revisit the concept of using shading primitives to support object class detection. We demonstrate and discuss the applicability of this approach to real world images of a standard benchmark data set. Experimental results suggest that our shading cues can be useful for object class detection.