Global Dilated Attention and Target Focusing Network for Robust Tracking

Self Attention has shown the excellent performance in tracking due to its global modeling capability. However, it brings two challenges: First, receptive field less attention on local structure and inter-channel associations, which limits semantics distinguish objects backgrounds; Second, feature fusion with linear process cannot avoid interference of non-target semantic objects. To solve above issues, this paper proposes a robust method named GdaTFT by defining Global Dilated (GDA) Target Focusing Network (TFN). The GDA provides new approach enhance while eliminating background. It is defined via focusing module, dilated channel adaption module. Thus, promotes key information, building long-range dependencies enhancing channels. Subsequently, target both rich semantics, TFN proposed accurately focus region. Different from present fusion, uses template as query build point-to-point correlation between search region, finally achieves part-level augmentation efficiently augments embedding weakening Experiments challenging benchmarks (LaSOT, TrackingNet, GOT-10k, OTB-100) demonstrate that outperforms many state-of-the-art trackers leading performance. Code will be available.