Existing tracking methods might suffer from the performance degradation due to insufficient positive samples. A typical network structure is proposed to enrich positive samples by generating masks during the tracking process. Although this structure has achieved good results, it ignores the drift problem that occurs when the tracked object is very similar to the surrounding objects. This problem is particularly significant when background interference exists and similar objects appear. To handle this problem, in this paper, we propose a novel attentive adversarial network for visual tracking. Inspired by human visual cognitive system, we propose to employ an attention mechanism to focus on each region differing the target object from the background. At the same time, we use a variant of the cross entropy (CE) function to deal with the class imbalance problem. Our network shows favorable performance compared with state-of-the-art methods on existing tracking benchmark datasets. We conclude that our novel attentive adversarial network not only enriches positive samples in the feature space but also prevents the similarity drift problem.

现有的跟踪方法可能由于正样本不足而导致性能下降。提出了一种典型的网络结构，通过在跟踪过程中生成掩码来丰富正样本。尽管此结构取得了良好的效果，但它忽略了当被跟踪对象与周围对象非常相似时发生的漂移问题。当存在背景干扰并且出现类似物体时，此问题特别重要。为了解决这个问题，在本文中，我们提出了一种新颖的注意力对抗网络，用于视觉跟踪。受人类视觉认知系统的启发，我们建议采用一种注意力机制来关注目标对象与背景不同的每个区域。同时，我们使用交叉熵（CE）函数的变体来处理类不平衡问题。与现有跟踪基准数据集上的最新方法相比，我们的网络显示出良好的性能。我们得出的结论是，我们新颖的专心对抗网络不仅可以丰富特征空间中的正样本，而且还可以防止相似性漂移问题。