REVT: Robust and Efficient Visual Tracking by Region-Convolutional Regression Network

This paper proposes a novel approach, namely REVT, for visual tracking based on a region convolutional regression network. REVT runs according to a coarse to fine scheme. It first builds an on-line update deep network to roughly select a candidate region in a fast way. It then refines the result by exquisitely searching the target within the candidate region by a deep regression network, which is trained off-line to account for more diverse intra-class appearance changes. REVT thus integrates the advantages of the two types of deep models, and demonstrates a good trade-off between accuracy and efficiency. We perform extensive experiments on the OTB-2013 and OTB-2015 benchmarks, and REVT reports competitive performance at a speed of 19 fps, proving its competency.