Adaptive Adversarial Example Generating Network for Object Detection

The consensus of most recent researches is that rich training data which includes more different object forms in different situations is beneficial to improve the performance of deep networks. Deep learning is very dependent on data diversity. So, we provide a novel and principle method called adaptive adversarial example generating network (AAEGN) to generate examples with occlusions for object detection. AAEGN can generate occlusion examples which are rare in current dataset and are difficult to localize and classify. Compared to the traditional adversarial network for the example generation, AAEGN employs several innovations to simplify the architecture of adversarial network and improve the example generation and training speed. What’s more, AAEGN can adjust the hyper-parameters of the adversarial network in the process of training adaptively. The results of our experiments show that using AAEGN to train Fast-RCNN pipeline can get a 1.8% boost on VOC2007 dataset and 3.8% boost on VOC2012 dataset.