3D pose refinement using rendering and texture-based matching

This paper presents a method for accurately determining the pose of Lambertian rigid objects present in an image. An initial pose estimate computed with the aid of local point features is ameliorated by considering all visible object texture. This is achieved by combining a textured mesh model of the object with a graphics renderer to synthesize an image of the object as would be captured by the camera at a particular pose. A rendered image is compared against the acquired one with the aid of a visual dissimilarity score involving cross-correlation. Populationbased stochastic optimization is used to efficiently search the pose space and minimize the dissimilarity between rendered images corresponding to candidate poses and the acquired image. The method is demonstrated with the aid of real and synthetic images. © 2014 Springer International Publishing Switzerland.