Real-Time Visual Tracking Base on SiamRPN with Generalized Intersection over Union

Abstract

Deep learning-based tracking methods have great challenges to handle larger training data with aiming to be invariant to all sorts of appearance variations. In this paper, we incorporate a novel Generalized intersection over union (GIOU) as bounding box regression loss into Siamese framework based tracker, and propose a visual tracking method based on Siamese region proposal network (SiamRPN) with generalized intersection over union. We set out to bridge the gap between optimizing the commonly used bounding box regression loss and maximizing the Intersection over Union (IOU) metric value. Our target estimation component is trained to predict the overlap between the target object and an estimated bounding box. Moreover, it can relieve the case that non-overlapping bounding boxes in training phase. Experimental validations have shown that our tracker performs substantially improvement on the tracking benchmarks OTB100 and is effective to deformation, occlusion and other challenges in object tracking.