Applying spatial restrictions to part-based detection and tracking
- Autores: Brais Martínez
- Directores de la Tesis: Xavier Binefa i Valls (dir. tes.)
, Ricardo Juan Toledo Morales (dir. tes.) - Lectura: En la Universitat Autònoma de Barcelona ( España ) en 2010
- Idioma: inglés
- Tribunal Calificador de la Tesis: Jordi Vitrià Marca (presid.) , Jordi González Sabaté (secret.) , Maja Pantic (voc.)
- Enlaces
- Tesis en acceso abierto en: TESEO
- Resumen
The problem tackled in this thesis relies on the definition of objects depending on a set of its constituent parts. The advantages of this definition are several. Among them, we emphasize its relevance to the ease of modelling target aspect variations in the presence of complex target movements, as out-of-plane rotations or flexible movements. Furthermore, when dealing with class-specific tasks, the intra-category differences of the different instances can be reduced with such a representation. However, this powerful definition opens new questions, as how to define the spatial relations between the different parts of an object in order to get a good trade-off between representation power and flexibility.
This thesis provides several possible solutions, adapted to different scenarios. The techniques used in here range from purely computer vision techniques to a combination where machine learning is of large importance. Also, the geometric content of the ideas described is constant throughout the thesis, so both the problem formulation and the solutions provided have an important intuitive interpretation.
