Texturing 3D Models with Low Geometric Features

This paper presents a texture projection and alignment method specifically tailored for objects with low geometric features. A common way to build virtual replicas is to acquire the geometry with a 3D scanner and model its reflectance properties (i.e. colors) by projecting photos onto the 3D surface. To correctly align each photo it is necessary to retrieve the camera´s extrinsic matrix, which is usually achieved by optimization algorithms that match geometric features of the model with their corresponding ones on the photos. However, when dealing with cultural heritage artifacts or mechanical pieces, for example, some models inherently lack significant geometric detail. We propose a method to automatically extract these features from high resolution photos and register them in a sequential manner using a variation of the contour-based approaches. The algorithm requires minimum user intervention as only an initial guess of the camera´s direction is required for each photo. In addition, we describe an improved formulation of the optimization method based on a smooth function that avoids typical local minima found in this scenario. The results are illustrated with high-resolution textured models produced by our system.