In this thesis, we propose new efficient methods for the multiple object segmentation task based on a hierarchical approach of layered graphs. As input, we have an image high-level priors of each object, different local image constraints and high-level priors for each object. Then, we create a hierarchical graph where each level is also a graph, and setting structural constraints among levels, a single energy optimization is performed leading to globally optimal results. Compared other multi-object segmentation methods, our approach is less restrictive, leading to globally optimal results in more general scenarios such as in medical and synthetic images in 2D and 3D, and using superpixels instead of pixels.
This project involved the study and implementation of different matrix factorization techniques for recommender systems.
This thesis explores the problem of object detection in videos using models created from a set of static images. The model was given by a Mixture of Deformable Parts. We applied those models to classify objects in video sequences obtained from static or dynamic cameras. The test scene was a totally different one than the used in the training process. My master was funded by FAPESP. During my master's, I have participated in the summer school of INRIA (2012) with a poster presentation and in SIBGRAPI conference (2012) where has been invited to write an extension work to the Pattern Recognition Letters which was accepted for publication. I got the second prize in the XX Latin American Contest of Master Thesis (CLTM - CLEI 2013).
We performed experiments using the SIFT technique to tag some instance of objects on different images of the same instance but in different viewpoints.