An Extended Geometric Approach in Active Vision to Detect and Eliminate Specularity in Shiny Colored Objects

In this paper we present an extended geometric approach, that detects and eliminate the specular points in a texture image of shiny objects, with the help of corresponding range data. The approach neither count on spectral variations of the texture image nor employs rigid constrains on illumination sources, such as point light source limitations. It rather determines the candidate specular points, based on range data measured at the same view direction as the texture image, the viewing geometry, the geometry of the illumination source, and provide spatial information of such points to the secondary processing algorithms. The specular elimination process is embedded with a viewpoint shift algorithm, which uses an active vision methodology, based on data supplied by the geometric method. Construction of realistic 3-D models of objects has been sought in a variety of computer vision application areas ranging from computer graphics, CAD, virtual reality, electronic catalogues, and so on. 3-D model acquisition has been extensively discussed by the computer vision research community, and a number of approaches have been proposed in the past [I]. In this work we adopt an active vision oriented approach for 3-D model acquisition, which could be able to construct models of real world complex objects. Such a system has previously been demonstrated in [2], where a rangefindercalled the "Cubicscope" is mounted on a articular robot arm with five degrees of freedom. It hasbeen noted however, that the constructed model possess with some undesirable texture variations due to the specularity in corresponding texture images. There are a number of approaches proposed by various researches to deal with highlights in texture images upto date. Most of such methods however have their own limitations or restrictions when applied to general cases. The dichromatic reflection model, proposed by Shafer el. at. [3], bears the application limitation for textures with almost similar spectral variations as the illumination source. In such cases, body and surface reflection clusters would almost be coincide, thus making it hard to separate. Other well known reflectionmodels, such as Beckman, and Spizzichino [4], Cook and Torrence [5], either contribute to the limitation of point light source restrictions or require extensive computations to track down specular spots in a texture image. Such limitations would adversely affect to the needs of our system, which encourages a technique that could be implemented as an intermediate step to the said 3-D modeling task. Therefore the core of this study is devoted to finding a robust yet simply implementable highlight extraction method, which will be well suited as an intermediate task in an active vision oriented 3-D object modeling activity. As described in the previous section, the consideration is mainly based on constructing a simple, yet a robust technique, that extract highlight components, which can be implemented as an intermediate processing step of the 3-D modeling system. Let us consider the geometry of highlight formation in a texture image, captured from a pin-hole camera (fig. 1).