Enhancing object detection in mobile augmented reality: A novel framework integrating knowledge distillation and unsupervised domain adaptation

Abstract

Object detection plays a crucial role in enhancing mobile augmented reality (MAR) applications, but the computational limitations of mobile devices and the dynamic real-world environments pose significant challenges. This study proposes a novel framework that integrates Knowledge Distillation (KD) and Unsupervised Domain Adaptation (UDA) to address these issues. KD transfers knowledge from a resource-heavy "teacher" model to a lightweight "student" model optimized for mobile deployment, while UDA enables the student model to adapt to real-world conditions without labeled data. Our framework uses YOLOv5 models, where the student model, YOLOv5n, learns from the teacher model, YOLOv5 small, improving precision and maintaining efficiency. Experiments on the VOC2007 and COCO datasets show that our SKD-UDA net achieves 78.2% mAP at IoU 0.5 and 50.8% mAP at IoU 0.5:0.95, outperforming the baseline YOLOv5n by 5.5% and 5.7%, respectively, without increasing the model size (1.9 MB). This approach enhances accuracy and computational efficiency, making it ideal for MAR applications. Our contributions advance object detection in MAR, improving user interaction by increasing detection accuracy, inference speed, and seamless integration of digital and physical environments.