Understanding Siamese Networks for Visual Object Tracking

Abstract

Visual tracking of an arbitrary object has drawn great attention in the computer vision community for years. Visual object tracking methods based on Siamese networks achieved state-of-the-art results with balanced accuracy and speed in the past few years. Understanding those trackers may shed light on future research on this topic, which makes it no less important than just incrementally improving the performance of the trackers. In this report, we give a comprehensive analysis of multiple representative trackers based on Siamese Networks. Specifically, we depict the developments of Siamese networks in visual object tracking from the earliest one Siamese-FC to the latest ones. Besides, we give both quantitative and qualitative analysis of the performance and discrimination power of Siamese networks based trackers, by evaluating them on several canonical tracking benchmarks including VOT and OTB datasets and visualizing feature maps. We also identify several causes of failures. Some future research directions are proposed in the end.