De-Background Generative Adversarial Networks for Object Detection

De-Background Generative Adversarial Networks for Object Detection

연구 분야: Artificial Intelligence

학회: CIPAE 2020: Proceedings of the 2020 International Conference on Computers, Information Processing and Advanced Education

Region proposal algorithms tend to become the mainstream among most object detection tasks. As a prominent representative, Faster R-CNN introduces region proposal networks (RPN), a deep neural network, to tell the unified network where to put more attention on. Based on the huge success of deep neural network, we are convinced that Generative Adversarial Networks (GAN) can also be competent for this task. For this purpose, we introduce a novel use of GAN, called De-background Generative Adversarial Networks (DBGAN), to generate the bounding boxes in one image under the simple background, namely detect object locations. We use intersection-overlap-union (IoU) to measure the quality of the generated boxes and we get good results on PASCAL VOC 2007, 2012 datasets.