Perceptual information-theoretic measures for viewpoint selection and object recognition
Xavier Bonaventura Brugués · TDX (Tesis Doctorals en Xarxa) · 2015
Viewpoint selection has been an emerging area in computer graphics for some years, and it is now getting maturity with applications in fields such as scene navigation, volume visualization, object recognition, mesh simplification, and camera placement. But why is viewpoint selection important? For instance, automated viewpoint selection could play an important role when selecting a representative model by exploring a large 3D model database in as little time as possible. Such an application could show the model view that allows for ready recognition or understanding of the underlying 3D model. An ideal view should strive to capture the maximum information of the 3D model, such as its main characteristics, parts, functionalities, etc. The quality of this view could affect the number of models that the artist can explore in a certain period of time. In this thesis, we present an information-theoretic framework for viewpoint selection and object recognition. From a visibility channel between a set of viewpoints and the polygons of a 3D model we obtain several viewpoint quality measures from the respective decompositions of mutual information. We also review and compare in a common framework the most relevant viewpoint quality measures for polygonal models presented in the literature. From the information associated to the polygons of a model, we obtain several shading approaches to improve the object recognition and the shape perception. We also use this polygonal information to select the best views of a 3D model and to explore it. We use these polygonal information measures to enhance the visualization of a 3D terrain model generated from textured geometry coming from real data. Finally, we analyze the application of the viewpoint quality measures presented in this thesis to compute the shape similarity between 3D polygonal models. The information of the set of viewpoints is seen as a shape descriptor of the model. Then, given two models, their similarity is obtained by performing a registration process between the corresponding set of viewpoints La seleccio de punts de vista ha estat una area emergent en la computacio grafica des de fa alguns anys i ara esta aconseguint la maduresa amb aplicacions en camps com la navegacio d’una escena, la visualitzacio de volums, el reconeixement d’objectes, la simplificacio d’una malla i la col·locacio de la camera. Pero per que es important la seleccio del punt de vista? Per exemple, la automatitzacio de la seleccio de punts de vista podria tenir un paper important a l’hora de seleccionar un model representatiu mitjancant l’exploracio d’una gran base de dades de models 3D en el menor temps possible. Aquesta aplicacio podria mostrar la vista del model que permet el millor reconeixement o comprensio del model 3D. Un punt de vista ideal ha de captar la maxima informacio del model 3D, com per exemple les seves principals caracteristiques, parts, funcionalitats, etc. La qualitat d’aquest punt de vista pot afectar el nombre de models que l’artista pot explorar en un determinat periode de temps. En aquesta tesi, es presenta un marc de teoria de la informacio per a la seleccio de punts de vista i el reconeixement d’objectes. Obtenim diverses mesures de qualitat de punt de vista a traves de la descomposicio de la informacio mutua d’un canal de visibilitat entre un conjunt de punts de vista i els poligons d’un model 3D. Tambe revisem i comparem en un marc comu les mesures mes rellevants que s’han presentat a la literatura sobre la qualitat d’un punt de vista d’un model poligonal. A partir de la informacio associada als poligons d’un model, obtenim diversos tipus de renderitzat per millorar el reconeixement d’objectes i la percepcio de la forma. Utilitzem aquesta informacio poligonal per seleccionar les millors vistes d’un model 3D i per la seva exploracio. Tambe usem aquestes mesures d’informacio poligonal per millorar la visualitzacio d’un model de terreny 3D amb textures generat a partir de dades reals. Finalment, s’analitza l’aplicacio de les mesures de qualitat de punt de vista presentades en aquesta tesi per calcular la similitud entre dos models poligonals. La informacio del conjunt de punts de vista es vista com un descriptor del model. Llavors, donats dos models poligonals, la seva similitud s’obte mitjancant la realitzacio d’un proces de registre entre els conjunts de punts de vista corresponents