A region-centered topic model for object discovery and category-based image segmentation

Abstract

Latent topic models have become a popular paradigm in many computer vision applications due to their capability to unsupervisely discover semantics in visual content. Relying on the Bag-of-Words representation, they consider images as mixtures of latent topics that generate visual words according to some specific distributions. However, the performance of these methods is still limited by the way in which they take into account the spatial distribution of visual words and, what is even more important, the currently used appearance distributions. In this paper, we propose a novel region-centered latent topic model that introduces two main contributions: first, an improved spatial context model that allows for considering inter-topic inter-region influences; and second, an advanced region-based appearance distribution built on the Kernel Logistic Regressor. It is worth highlighting that the proposed contributions have been seamlessly integrated in the model, so that all the parameters are concurrently estimated using a unified inference process. Furthermore, the proposed model has been extended to work in both unsupervised and supervised modes. Our results for unsupervised mode improve 30% those of previous latent topic models. For supervised mode, where discriminative approaches are preponderant, our results are quite close to those of discriminative state-of-the-art methods.