Multi-view coding of local features in visual sensor networks

Local visual features extracted from multiple camera views are employed nowadays in several application scenarios, such as object recognition, disparity matching, image stitching and many others. In several cases, local features need to be transmitted or stored on resource-limited devices, thus calling for efficient coding techniques. While recent works have addressed the problem of efficiently compressing local features extracted from still images or video sequences, in this paper we propose and evaluate an architecture for coding features extracted from multiple, overlapping views. The proposed Multi-View Feature Coding architecture can be applied to either real-valued or binary features, and allows to obtain bitrate reductions in the order of 10-20% with respect to simulcast coding.