Semi-supervised Visual Tracking Based on Variational Siamese Network

Abstract

AbstractFor object tracking in a visual dynamic data-driven application systems (DDDAS) framework, visual appearance features can be extracted by the convolutional neural network, which has been shown to provide a robust feature representation. In this paper, a new semi-supervised learning framework, variational Siamese neural network, is developed for visual tracking by combining a Siamese network with a variational autoencoder, which supports both supervised and unsupervised training. The learned features are represented as Gaussian distributions in feature space, and the object is represented as a distribution in image space. The similarity between objects’ features is measured by an information theoretic distance. The tracking algorithm is based on the detection network’s detections to update the object state estimate. Experiment results show that the proposed visual tracking framework outperforms existing state of the art visual tracking approaches. KeywordsObject trackingSiamese networkSemi-supervised learning.