Creating and Recognizing 3D Objects

This thesis aims at investigating on 3D Computer Vision, a research topic which is gathering even increasing attention thanks to the more and more widespread availability of affordable 3D visual sensor, such as, in particular consumer grade RGB-D cameras. The contribution of this dissertation is twofold. First, the work addresses how to compactly represent the content of images acquired with RGB-D cameras. Second, the thesis focuses on 3D Reconstruction, key issue to efficiently populate the databases of 3D models deployed in object/category recognition scenarios. As 3D Registration plays a fundamental role in 3D Reconstruction, the former part of the thesis proposes a pipeline for coarse registration of point clouds that is entirely based on the computation of 3D Local Reference Frames (LRF). Unlike related work in literature, we also propose a comprehensive experimental evaluation based on diverse kinds of data (such as those acquired by laser scanners, RGB-D and stereo cameras) as well as on quantitative comparison with respect to three other methods. Driven by the ever-lower costs and growing distribution of 3D sensing devices, we expect broad-scale integration of depth sensing into mobile devices to be forthcoming. Accordingly, the thesis investigates on merging appearance and shape information for Mobile Visual Search and focuses on encoding RGB-D images in compact binary codes. An extensive experimental analysis on three state-of-the-art datasets, addressing both category and instance recognition scenarios, has led to the development of an RGB-D search engine architecture that can attain high recognition rates with peculiarly moderate bandwidth requirements. Our experiments also include a comparison with the CDVS (Compact Descriptors for Visual Search) pipeline, candidate to become part of the MPEG-7 standard.