Recuperação de propriedades fotométricas de cenas com iluminação global a partir de uma única imagem

Abstract

Obtaining the photometric properties of a scene considering it contains textured surfaces is a challenge on Computer Vision. In general, the existing algorithms will estimate the surface's photometric parameters in every scene's 3D point, which is a very expensive optimization process that also demands as input several images of the scene. Moreover,many of those methods does not consider reflection and transparency among the scene's objects. In this thesis, we present a methodology to obtain the photometric properties of textured surfaces from a single image of the scene, considering that light can reflect and refract among the scene's objects. The method assumes that the following informationabout the scene is known a priori: the 3D geometry; information on the position and intensity of the light sources; a single image; the refraction indices of the objects; a fixed and unique set of non-Lambertian photometric parameters for each surface. The solution of a linear equation system will describe the variation of the photometric propertiesof the surfaces observed directly or indirectly by the input image. Experimental results are shown using a real scene, and it demonstrates that the method is efficient and precise, as we can see by the resulting images its corresponding processing time.