Data Augmentation using GAN for Multi-Domain Network-based Human Tracking
Kexin Chen, Xue Song Zhou, Wei Xiang, Qidong Zhou · 2018
This paper presents an on-line data augmentation method for discriminative Convolutional Neural Network (CNN)-based human tracking. Different from randomly sampling around the object, we propose a novel hard negative mining method based on Generative Adversarial Networks (GAN). In order to increase distraction and decrease the redundancy of negatives, the samples with similar appearance with positives generated by GAN generator are treated as hard negatives. Moreover, we integrate this hard negative mining with on-line updating mechanism into multi-domain network (MDNet)-based tracking framework, which makes the network become more discriminative as the learning proceeds. Experimental results on existing tracking benchmark demonstrate the effectiveness and robustness of our proposed method, especially for the long time tracking.