Single view single light multispectral object segmentation

Abstract

In this paper we present an approach for the acquisition and segmentation of spectral Bidirectional Reflectance Distribution Function (BRDF) measurements of real-world objects. The acquisition setup is a priori fully calibrated and provides pixel-synchronous image and depth data of the examined objects. Based on one single viewing and illumination geometry, we are able to determine spectrally distinct surface regions for objects with abruptly changing surface materials (painted surface patches) and for objects with gradually changing materials (partially oxidized iron). For clustering we apply the k-means algorithm and the mean-shift algorithm. The segmented clusters are used to adapt individual spectral BRDFs (Lambert, Phong, Cook-Torrance) to the obtained cluster data. Additionally, the elemental abundances of iron and rust on a metal surface are analyzed using spectral unmixing. The paper presents a detailed discussion of our method and provides critical insight into the obtained results.