Augmented Reality for Real Object Detection

Abstract

In the era of AI booming, object detection is an essential technology for computer vision tasks and widely adopted in autonomous driving. We propose a method to enhance object detection accuracy by adding virtual objects to real scenes through augmented reality, thereby quickly generating a large amount of data to facilitate model training. In addition, Augmented Reality (AR) can create data for rare scenarios in real worlds, such as a car flipping over on the road or a cargo overturned, which can alleviate the long-tail problem of AI models. Furthermore, our tool can generate both 2D and 3D bounding boxes directly. To verify our method, we performed transfer learning on YOLOv7 pre-trained model using 30,766 AR synthesized images of 4 traffic-related classes: Person, Car, Bicycle and Motorcycle. The new detector was evaluated on the COCO dataset. Experiments showed that our method can increase the detector accuracy as well its ability of detecting small objects.