Adversarial Learning-based Data Augmentation for Rotation-robust Human Tracking

This paper analyzes the diversity deficiency of positive training samples used to fine-tune CNN-based tracking networks, especially when confronted with large pose changes and out-of-plane rotation challenges. Therefore, we present a novel adversarial learning-based hard positives generation method and embed it into the multi-domain network (MDNet)-based tracking framework. Instead of adopting the dense sampling strategy to generate monotonous positive samples, we cast it as a cross-domain image transformation problem, which is designed to be able to generate hard positive samples with more diversity and some degree of motion blur and pose direction changes. Experimental results on tracking benchmark demonstrate the effectiveness and robustness of our proposed method compared with the state-of-art trackers.