Geometry and appearance modelling from images

Non disponibile

This thesis is about the acquisition of a textured 3D surface model from a sequence of sparse images, in the direction of a completely un- supervised approach. It offers original contributions in the following fields. Autocalibration. The problem of autocalibration of a moving cam- era with unknown constant intrinsic parameters is here addressed. Existing autocalibration techniques use numerical optimization al- gorithms whose convergence to the correct result cannot be guaran- teed, in general. To face this problem, a method where an interval branch-and-bound method is employed for numerical minimization has been deveoped. Thanks to the properties of Interval Analysis this method converges to the global solution with mathematical certainty and arbitrary accuracy, and the only input information it requires from the user are a set of point correspondences and a search interval. Triangulation. We deal with the problem of an accurate estimation of 3D points coordinates that rigorously takes into account the propa- gation of data errors and roundoff. Image points are represented as small rectangles: as a result, the output of the n-views triangulation is not a single point in space, but a polyhedron that contains all the possible solutions. Computational Geometry techniques are used to estimate this polyhedron. Constraint-based reconstruction. We address the problem of un- supervised constrained scene modelling from many calibrated views. Given connectivity information, planes are detected from an ap- proximate model and geometric constraints are imposed on such planes. Then, these constraints are used as well with the polyhe- dral bounds for 3D points to obtain a faithful 3D model. All this process is completely automatic, with a further step of constraints simplification, aiming at eliminating redundancies and erroneous contraints. Reflectance recovery. The problem of eliminating lighting artefacts from the input images of the 3D model, obtaining normalized tex- ture maps, which can be re-illuminated, is here addressed. Assum- ing Lambertian surfaces, we separate reflectance information from shading, forming two separate images. The approach is based on the binary classification of image derivatives, with the help of the grayscale image, invariant to illumination, suggested by Finlayson et al. The results obtained are comparable with the state of the art, but they are obtained without any learning processes to classify derivatives.