Active Object Detection On Graphs Via Locally Informative Trees
Daphney-Stavroula Zois, University of Illinois, Urbana-Champaign
Maxim Raginsky, University of Illinois, Urbana-Champaign

Abstract:
Active object detection refers to the problem of determining the existence and location of objects in an image by actively selecting which regions of the image to explore. Herein, an object detection algorithm is proposed that models image regions as vertices and overlap relationships as edges in a directed weighted graph. Information is propagated from labeled vertices through graph edges that operate as noisy channels via message passing over locally informative trees that are extracted from the original graph using an information–-theoretic criterion. Influential vertices are determined by an appropriate centrality index. Our algorithm can be applied on top of any state-of-the-art region proposal method as it treats it as a black box. The effectiveness of the proposed algorithm is illustrated on different scenarios, where in some cases only 0.45% of the total regions is evaluated.