An Improved Visual Tracking Algorithm Based on Efficient Convolution Operators

Effcient Convolution Operators (ECO) is one of the most outstanding visual tracking algorithms in recent years, it combines two methods of Deep Learning and Discriminative Correlation Filter (DCF), and has excellent performance in VOT2016, UAV123, OTB-2015 and TempleColor. The paper propose integrating ECO and Fully Convolutional Networks (FCN) to achieve state-of-the-art segmentation for ECO. In our experiments, the original CNN model of ECO is replaced by the FCN model. Compared with the traditional Convolutional Neural Networks (CNN), the FCN has higher accuracy of segmentation and can input any size of image. We perform comprehensive experiments which obtained 0.653 area under curve (AUC) and 0.861 precision plot (DP) on OTB-2013 dataset.