Synthesizing Training Data for Object Detection Using Generative Adversarial Networks

Abstract: Object detection is an important tool in computer vision and a popular application of machine learning. One of the main challenges in object detection, and machine learning in general, is acquiring sufficient training data. Many types of data can be hard or expensive to collect and label, or be subject to privacy concerns and regulations such as the General Data Protection Regulation (GDPR). This is particularly true for many object detection tasks, such as face detection where the training data consists of images depicting faces. Using synthetic data for training has been attempted before, but no consensus exists on how to best utilize it. This work focuses on using a priori trained Generative Adversarial Networks (GANs) to produce synthetic images of faces, and using them to train detectors based on Haar-like features. Experiments were conducted on both replacing real images with synthetic, and introducing synthetic variance by augmenting real images using image-to-image translating GANs. It was found that GAN-generated images can indeed be useful for detector training. Although real images consistently performs better, the amount of data plays a role as well, and a priori trained GANs can easily produce a lot of synthetic data with good variation. If real data is hard to collect, synthetic data produced by a GAN could be a viable option. It was also found that image-to-image translating GANs can be useful for data augmentation, especially when real data is scarce. Future work should focus on the impact of variance and bias in the synthetic data and how it can be controlled for optimal performance.