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

Highlights：

• We propose a label smoothing loss regularization to integrate unlabeled GAN-generated data with real labeled training data for classifier training, which introduces more color, lighting and pose variances to regularize the model and avoid model overfitting.

• We conservatively learn a reliable re-detection correlation filter, which is not only combined with classification score metric to evaluate tracking results, but also used to recover tracking failures.

• We extensively validate our method on five large-scale benchmark datasets: OTB-2013, OTB-100, UAV123, UAV20L, and VOT2016. Extensive experimental results demonstrate that the proposed tracking algorithm performs favorably against the state-of-the-art trackers.

摘要

•We propose two types of generative adversarial networks (GANs) to augment training data in the sample space and feature space respectively, which can capture a variety of appearance changes and bridge the gap between data hunger deep neural networks and visual tracking task.•We propose a label smoothing loss regularization to integrate unlabeled GAN-generated data with real labeled training data for classifier training, which introduces more color, lighting and pose variances to regularize the model and avoid model overfitting.•We conservatively learn a reliable re-detection correlation filter, which is not only combined with classification score metric to evaluate tracking results, but also used to recover tracking failures.•We extensively validate our method on five large-scale benchmark datasets: OTB-2013, OTB-100, UAV123, UAV20L, and VOT2016. Extensive experimental results demonstrate that the proposed tracking algorithm performs favorably against the state-of-the-art trackers.