Graph search and its application to image matching

• Autores: Jorge Artieda Trigueros
• Directores de la Tesis: Pablo San Segundo Carrillo (dir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2017
• Idioma: español
• Tribunal Calificador de la Tesis: Agustín Jiménez Avelló (presid.), Paloma de la Puente (secret.), Arturo de la Escalera Hueso (voc.), Ivana Ljubić (voc.), Fabio Furini (voc.)
• Programa de doctorado: Programa de Doctorado en Automática y Robótica por la Universidad Politécnica de Madrid
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • The main innovation presented in this thesis is a novel filtering technique for visual feature matching that takes into account multiple hypothesis. This algorithm, named CCMM, uses graph techniques to efficiently evaluate feasibility and joint compatibility of a set of possible matches. Specifically, the problem is reduced to a maximum clique search over an association graph and state-of-the-art algorithms are applied to solve it. Another contribution of this work is the tailoring of an exact clique solver for this application; the new algorithm is denoted BBMCW. BBMCW uses a bitstring formulation with sentinel bits to limit the boundaries of the semi-sparse bit-blocks that arise from the specific structure of the CCMM graphs.

    The CCMM algorithm is then applied to the Structure from Motion problem. There, it improves the reconstruction results and helps to maintain overall stability. In the context of the Structure from Motion problem, this thesis presents a further contribution concerning the initialization of the bundle adjustment optimization step. In the first place, it shows that initialization is critical for a successful reconstruction of the scene. Moreover, this work also describes several novel strategies for initialization and compares the scene reconstructions with those obtained by known strategies. The reported results show that ordering frames according to the minimum distance to a central point in an error graph improves state-of-the-art bundle adjustment.