Texturing 3D models from sequential photos

Abstract

This paper proposes a methodology to texture 3D objects with low geometric features from sequentially taken photos. These models pose a challenge to current approaches since they are mainly driven by geometric features—such as contours—that can be extracted from the photographs and uniquely matched with the 3D model. However, when dealing with certain types of objects, such as vases or mechanical equipments, for example, it is not uncommon to find cases where the geometric information is insufficient. Our method compensates for the lack of geometric features by using a variation of a contour-based approach that is guided not only by external contours, but also by the internal ones extracted directly from the photos. To align the features a custom-made optimization method is described that avoids common convergence pitfalls encountered in this scenario. In addition, pursuing a fully automatic solution, a linear approach based on feature matching is employed to generate a first guess for the nonlinear optimization. The overall goal is to facilitate an on-site registration process where the photos are taken in a sequential manner and aligned as they are acquired.