EFTrack: Enhanced fusion for visual object tracking

Abstract

Recently, deep learning-based networks for object tracking mainly adopt the single-stream single-stage framework. However, this approach often overlooks the backbone network’s own limitations. To address the issue, this paper utilizes an independent backbone network to directly construct the tracker and proposes optimizations. First, we propose a contour information enhancement (CIE) module to distinguish objects from the background through frequency domain filtering. Secondly, a patch information fusion (PIF) module is introduced to enable information interaction between non-overlapping patches. Furthermore, a lightweight multi-scale feature fusion module is proposed to enhance the backbone network’s capability to learn multi-scale information. The network’s generalization is enhanced using the DropMAE pre-trained model. The proposed tracker demonstrates superior performance on benchmark datasets, surpassing TATrack-B and SeqTrack-B384 networks by 3.4 % and 1.9 % respectively in terms of the AO metric on the GOT-10k dataset. The code is released at https://github.com/ Nirvanalll/EFTrack.