Learning to match using siamese network for object tracking

Abstract

The general object tracking problem traditionally been tackled by modeling the object’s appearance. In this paper we consider object tracking as a similarity measurement problem. We focus on learning a matching mechanism with great generalization ability. We present a Siamese convolutional neural network as a matching function to perform object tracking. First we simply match the exemplary target in previous frame with the candidates in a new frame using cosine similarity and return the most similar one by the learnt matching function. Then we perform bounding box regression to refine the target position given by the network as the final result. Extensive experiments on real-world benchmark datasets validate the superior performance of our approach.