Search

This work will focus on studying different deep learning architectures for obtaining depth information from monocular RGB images. During this project, state-of-the-art deep learning models have been used to estimate depth maps from a monocular RGB image applying a teacher-student learning approach. This paradigm has been used in order to distillate the knowledge of high capacity deep neural networks into shallower ones to make inference faster for real-time applications. Some successful applications of this technique can be found both at natural language and computer vision applications.

En este trabajo de fin de grado se exploran distintas técnicas de aprendizaje profundo (deep learning) con el objetivo de obtener una segmentación de imagen basada en criterios de uniformidad. Estudiamos varios algoritmos de segmentación aplicándolos a images de profundidad y transferimos la segmentación obtenida a la Pyramid Scene Parsing Network contando solo con la información de los canales RGB como entrada. This bachelors thesis explores different deep learning techniques to achieve an image segmentation based on uniform criteria. We study various state-of-art segmentation algorithms applying them on depth images. Moreover, we transfer the depth segmentation knowledge to Pyramid Scene Parsing Network using only RGB information as input. Aquest treball de fi de grau explora diferents tècniques d'aprenentatge profund (Deep Learning) per tal d'obtenir una segmentació d'imatge basada en criteris d'uniformitat. Estudiem varis algoritmes de segmentació ja existents aplicant-los a imatges de profunditat. Un cop fet això transferim aquest coneixement a la Pyramid Scene Parsing Network utilitzant tan sols els canals RGB com a entrada.