Department of Pattern Recognition

(eng): Material recognition applications use typically color texture-based features; however, the underlying measurements are in several application fields unavailable or too expensive. Therefore, bidirectional reflectance measurements are used, i.e., dependent on both illumination and viewing directions. But even measurement of such BRDF data is very time- and resources-demanding. In this paper we use dependency-aware feature selection method to identify very sparse set of the most discriminative bidirectional reflectance samples that can reliably distinguish between three types of materials from BRDF database - fabric, wood, and leather. We conclude that ten gray-scale samples primarily at high illumination and viewing elevations are sufficient to identify type of material with accuracy over 96/%. We analyze estimated placement of the bidirectional samples for discrimination between different types of materials. The stability of such directional samples is very high as was verified by an additional leave-one-out classification experiment.