TA-CFNet: A New CFNet with Target Aware for Object Tracking

摘要

A new network constructed by the combination of Siamese network and correlation filter has achieved enormous popularity because Siamese networks obtain high accuracy and correlation filters provide amazing speed. How to tackle boundary effects problems brought by filters, to fuse CF layers with multi-layers of CNN is an essential problem of the new network. Most papers deal with the problem by simply adding cosine window to every image. However, if the target is too small for bounding box to include background information or if target is too big for bounding box to loss partial information. In this paper, a new CFNet with target aware (TA-CFNet) for object tracking is proposed. TA-CFNet intergrates current target position and feature weight map to form target likelihood matrix. This target likelihood matrix is used to optimize and update the correlation filter, so that the template object of the deep tracking network is framed as accurately as possible. Experimental results on OTB benchmarks for visual tracking demonstrate that our proposed method outperforms other trackers in deep learning.