Robust Visual Tracking based on Adversarial Unlabeled Instance Generation with Label Smoothing Loss Regularization

Abstract

Recent studies have shown that deep neural networks have pushed visual tracking accuracy to new heights, but finding more robust long-term tracking is still challenging because of the dynamic foreground and background changes. This phenomenon affects the overall performance via online training sample generation. The dense sampling strategy has been widely used for its convenience, the appearance variation is severely limited by its highly spatial overlapping mechanism. The sample candidate evaluation with a classification score metric is not always reliable throughout the entire process, therefore, tracking failure is inevitable. As an effective solution, this paper proposes a novel sample-level generative adversarial network (GAN) to enrich the training data by generating massive amounts of sample-level GAN samples. These samples are not only similar to the real-life scenarios, but also could carry more diversity of deformation and motion blur to a certain degree. For occlusion invariance, a feature-level GAN is incorporated to generate more challenging feature-level GAN data by creating random occlusion masks in deep feature space. To facilitate the online learning process, a label smoothing loss regularization is introduced to achieve model regularization and over-fitting reduction by integrating the unlabeled GAN-generated training data with the realistically labeled ones. In addition, a re-detection correlation filter conservatively trained with reliable training data is employed to integrate a classification score metric to perform reliable model updates and avoid heavy degradation. Furthermore, we also carry out the re-detection correlation filter on the candidate region proposals to handle the tracking failures. The proposed tracker has shown superior performance in comparison to the other state-of-the-art tracking approaches on the OTB-2013, OTB-100, UAV123, UAV20L, and VOT2016 benchmark datasets.