Towards automatic acquisition of high-level 3D models from images

In recent years there has been a surge of interest in automatic modeling from images. While the current state of the art in three-dimensional reconstruction has focused on the recovery of dense and accurate representations of objects imaged through pictures or video, the sustained interest in accessible modeling software is a strong evidence of an untapped general need for compact, abstract representations of objects. In this thesis, the problem of producing high level models starting from images is discussed in details. In the first part, an automatic uncalibrated Structure from Motion pipeline is presented. Starting from the output of the pipeline, two different approaches of generating high-level renditions are studied. The first approach employs a novel Multiple view Stereo algorithm to produce a dense and accurate point cloud. A retrieval system for meshes, based on segmentation and Bag of Words, is then introduced. In the latter approach, the sparse Structure from Motion point cloud is fitted by planes and planar patches. Planar patches are a compact, intermediate representation of the scene. Both branches of the thesis aim to narrow the gap between scene acquisition and interpretation, through the definition of high level renditions produced by very different strategies.