IX Workshop de Visão Computacional (WVC 2013)

In several domains, learning a dictionary of features from data provides adaptive schemes which are possibly well suited for addressing reconstruction and classification tasks. Regularization methods provide an ideal setting for dictionary learning since the choice of the penalties can be driven by the prior knowledge on the specific problem. In this talk we will review regularization methods in the context of learning, discuss some recent developments in the case of dictionary learning and present some experimental results in the domain of computer vision, biomedical image analysis and, time permitting, computational biology.

Image datasets from multiple imaging modalities (e.g., medical, remote sensing, photography, video) are commonly available to support research, education, entertainment, and several other activities. As these datasets grow large, image annotation by keywords becomes a crucial task for effective information retrieval due to the semantic gap between visual features and user expectations. In this scenario, manual image annotation is infeasible and active learning strategies are paramount. These techniques must select a relatively small training set with the most informative images for manual annotation/correction, such that the assigned keywords can be automatically propagated to the remaining images in the dataset. However, can we get high accuracy fast and with minimal human interaction? In this talk, I will present a new approach for active learning which answers that question. This approach effectively reduces and organizes a learning set of images, and selects the most representative ones for human annotation. The methods are based on the optimum-path forest framework, used for fast image clustering and classification. The talk will be divided into four parts. First, the image annotation problem, related works, and the importance of dataset reduction and organization are introduced. Second, the optimum-path forest clustering is used to considerably reduce the dataset into a learning set of images. A minimum-spanning tree of the learning set is then exploited for image organization. Third, an optimum-path forest classifier is trained in a few iterations of active learning, while it selects images from the organized learning set for user annotation/correction. Finally, I will justify the choice for the optimum-path forest framework by presenting results which also include our active learning approach using other clustering and classification methods.

The estimation of correspondences between two or more images is one of the fundamental problems in computer vision. It is related to many interesting fields such as optical flow estimation, stereo reconstruction, image registration and motion-aware video processing. Since first approaches were developed in the late 70s variational methods, i.e. global optimization methods, have been among the most popular techniques for solving such tasks. In the last three decades this popularity triggered a tremendous amount of research in the field. Nowadays global optimization methods belong to those methods that are best understood and that offer the best performance in terms of accuracy. This talk gives an overview of the modeling and the algorithmic concepts for variational methods for correspondence problems in computer vision. Thereby we address four important aspects of estimating correspondence problems with variational methods: (i) First, it gives insights into the design of highly precise models that are among the most accurate approaches in the literature. This is done by discussing recent modeling concepts for motion estimation and embedding them into a general implementation framework. (ii) Secondly, it presents efficient numerical algorithms both on sequential and parallel architectures. Despite the complexity of the underlying optimization problems, these algorithms allow to compute dense motion fields of excellent quality in real-time. (iii) Furthermore, the talk shows the extensibility of the discussed models to other important correspondence problems in computer vision such as single and multi-view stereo reconstruction, scene flow, and facial performance capture. By jointly estimating the 3-D motion, the depth and the scene geometry, those methods are able to provide results that are robust and accurate at the same time. (iv) Finally, it comments on several applications for which motion information plays a key role. Such applications include video processing, image registration as well as audio-visual speech recognition.