HERNÁNDEZ-RODRÍGUEZ, CASTELÁN: IMPROVING PHOTOMETRIC CONSISTENCY 1 A method for improving consistency in photometric databases Felipe Hernández-Rodríguez felipe.hernandez@cinvestav.edu.mx Mario Castelán mario.castelan@cinvestav.edu.mx Robotics and Advanced Manufacturing Centro de Investigación y de Estudios Avanzados del I.P.N. Ramos Arizpe, Coah., México. Abstract Building photometric databases usually requires the gathering of images of a still object under different light source directions. During this process, unexpected artifacts such as noise, shadows, inter-reflections and other unwanted effects introduced by the sensibility of the camera may appear along the database, diminishing its consistency as a whole and therefore its suitability for the purposes of photometric analysis. This paper describes a method for improving photometric consistency in image databases acquired under photometric rigs. The main idea of our approach is to build and analyze a lumi- nance matrix storing the reflectance behavior of each pixel under the different light source directions. To this end, we propose to fit sinusoidal functions to the singular vectors of this luminance matrix in order to improve its agreement with Lambertian reflectance. Experiments demonstrate that our method improves the photometric consistency of the database, providing stability for the purposes of photometric analysis of the database and surface shape recovery. 1 Introduction Estimating 3D shape and reflectance from imagery is a relevant topic in computer vision and computer graphics, since it simplifies the task of modeling the appearance of objects. Techniques based on photometric changes, such as the photometric stereo method (PSM), have proved to be useful for such task by enforcing illumination variations over a still object. A quick review in the area reveals that issues such as the strategic placement of light sources [2, 4], the modeling of Bidirectional Reflectance Distribution Functions (BRDF) to handle non-Lambertian reflectance [8] or shaded regions [1, 7], and the statistical properties of photometric databases [6] have received major attention in the field. Similarly, approaches such as the photometric sampling [12] have also been developed in order to recover shape and reflectance properties through photometric features. The pho- tometric sampling consists of measuring the response of a still object while a light source, placed on a turntable, moves around a single circular path. Research efforts in photometric sampling have been mostly focused on the recovery of surface normals by sine fitting, where the key idea is to adjust a sinusoidal function to the pixel’s response along the light source trajectory. Sine fitting has shown a benefit in determining the azimuth angles of the surface normals as well as in determining specular areas [10, 13]. c  2012. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms.