Object detection based on generative adversarial networks

Abstract

Recently, object detection has presented superior performance using the deep convolutional neural network (CNN). However, most CNN-based models need the bounding box information of the input image in pairs. To overcome this limitation, we propose the Generative Object Detection which learns with only cropped images that are not in pairs. Our model based on Generative Adversarial Networks (GAN) creates cropped images by making a mask that represents the object region. To achieve this goal, we devise a novel mask mean loss (MML) that helps the GAN be able to estimate the distribution of training data and uses dilated convolution for a wider reception eld in the generator. The experimental results show that Generative Object Detection improves mIoU and accuracy.