Deformable Convolutional Networks Tracker

Abstract

Object tracking is a fundamental topic in computer vision. It is an interdisciplinary scientific filed involving machine learning, pattern recognition etc, and has a wide applicability. Although, convolutional neural networks (CNNs) have achieved significant success for visual recognition tasks. Deformation and scale variation of targets are huge challenges in object tracking. They cannot address these challenges for relying on massive amounts of data to build deep models. They are inherently limited to model geometric transformations due to the fixed geometry in the building blocks. In short, CNNs are inherently limited to model large and unknown transformation. In this paper, we introduced a deformable convolution module to construct a novel deformable convolutional networks tracker(DCT).The tracker illustrates outstanding performance in the Visual Tracker Benchmark (OTB)100[7] benchmark with scale variation and deformation attributes.