Offset-Adjustable Deformable Convolution and Region Proposal Network for Visual Tracking

Abstract

Spatial information is considered as the key information in many computer vision tasks including visual tracking. We attempt to use spatial information to get essential knowledge of the tracking targets in videos. Adapting to the geometry variations of the objects, the deformable convolutional networks (Deformable ConvNets) show the successful exploitation of the potential of spatial information. Through investigating the work of Deformable ConvNets, we propose the offset-adjustable deformable convolutional network for tracking. Provided with the offset adjustability technique, our network is able to have a good grasp of spatial information for the tracking objects. An appropriate target localization method is crucial to efficient tracking. One novel region proposal network with Siamese structure is designed to propose the candidate target regions and plays a vital role in promoting the effectiveness and efficiency of our model. We test our method on widely-applied OTB2015 and VOT2017 benchmark datasets. The experimental results show our method realizes state-of-the-art effects and achieves great improvement in comparison to other trackers.