DCFNETDISCRIMINANTCORRELATION.PDFVIP

DCFNET: DISCRIMINANT CORRELATION FILTERS NETWORK FOR VISUAL TRACKING Qiang Wang, Jin Gao, Junliang Xing, Mengdan Zhang, Weiming Hu CAS Center for Excellence in Brain Science and Intelligence Technology, National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences; University of Chinese Academy of Sciences ABSTRACT Discriminant Correlation Filters (DCF) based methods now become a kind of dominant approach to online object track- ing. The features used in these methods, however, are ei- ? ?(x) ther based on hand-crafted features like HoGs, or convolu- x : 125 125 3 125 125 D tional features trained independently from other tasks like im- × × × × DCF age classi?cation. In this work, we present an end-to-end lightweight network architecture, namely DCFNet, to learn ? ?(z) the convolutional features and perform the correlation track- z : 125 125 3 125 125 D ing process simultaneously. Speci?cally, we treat DCF as a × × × × special correlation ?lter layer added in a Siamese network, Fig. 1: The overall DCFNet architecture. and carefully derive the backpropagation through it by de?n- ing the network output as the probability heatmap of object speed by solving a ridge regression problem ef?ciently in location. Since the derivation is still carried out in Fourier fre- Fourier frequency domain. quency domain, the ef?ciency property of DCF is preserved. Tracking with DCF has been one of the biggest break- This enables our tracker to run at more than 60 FPS during test throughs in the community since the hand-crafted multi- time, while achieving a