Surgical Instrument Tracking for Capsulorhexis Eye Surgery Based on Siamese Networks

Siamese-based trackers have shown excellent performances in the field of visual object tracking. In most of these trackers, pre-defined anchor boxes are needed in order to precisely predict the scale and aspect ratio of a target which is a prohibitive task. In this paper, an effective visual tracker called SiamBAN (Siamese Box Adaptive Network) is used which exploits the expressive potency of the fully convolutional network (FCN). SiamBAN is a flexible framework since there is no necessity of the prior box design which leads in hyper-parameters avoidance. However, this framework cannot capture all of the template variations. To address this problem, another tracking framework for visual object tracking called Gradient-Guided Network (GradNet) is utilized which has a template update module. The two networks are implemented on the first version of ARAS-Farabi Tracking-based Capsulorhexis Dataset (ARFaTv1) which contains a number of videos related to Capsulorhexis surgery. The implementation results indicate that SiamBAN tracker has a superior efficiency than GradNet tracker in this specific task.