Improving GAN-based Domain Adaptation for Object Detection

Unlabeled or insufficiently labeled data is a common challenge in autonomous driving applications. Perception tasks like object detection in particular require annotated data, which is expensive to acquire, due to mostly manual labeling processes. One approach to saving label cost is using synthetic data and unsupervised domain adaptation to bridge the visual appearance gap towards the real application data. Knowing the placement of objects in rendered scenes, labels can be extracted automatically. However, lack of realism in rendered scenes leaves a domain gap between real and synthetic data. We transfer image-to-image adversarial domain adaptation methods from semantic segmentation to object detection, leveraging synthetic data for the task of vehicle detection in real images. This improved GAN-based image generation is combined with a standard object detector. This is in contrast to previous works, which also applied domain adaptation methods to the detector. With our adaptations we achieve state of the art results in domain adaption using generative adversarial networks on GTA to Cityscapes synthetic to real unsupervised domain transfer for vehicle detection.