Adaptive distractor-aware for siamese tracking via enhancement confidence evaluator

Siamese networks have recently attracted much attention due to balancing accuracy and speed in tracking. However, most Siamese trackers emphasize how to learn target features, ignoring background information. In this paper, a novel Siamese-based method via an adaptive distractor-aware and enhancement confidence evaluator is proposed. First, to make full use of the background information, a distractor model is designed, which is used to select the dominant feature channels of less interference for target representation. Additionally, to enable the target model to work well with challenging scenes, an adaptive weights-aware strategy is proposed, which can adjust the feature channel weights online according to the single-convolutional attention network with L2 loss. In addition, a dynamic template update strategy is proposed to adapt target appearance variations. The strategy evaluates the tracking results via a new correlation peak average energy (CPAE) evaluator, which has more confidence for good templates, guiding the fusion of reliable template samples and further enhancing the adaptability of the template and making the tracker more robust. Experimental results on widely used OTB100, TC-128, UAV123, VOT2016 and VOT2018 benchmarks verified that the proposed tracker achieved outstanding performance compared with state-of-the-art trackers.