A Novel Lightweight Deep-Learning Offloading Scheme for Try-On in 5G-Edge Network

The integration of Metaverse technology with Virtual Try-On (VTO) systems is revolutionizing the retail industry by offering immersive shopping experiences. This paper introduces Edge-VR4Fit, a novel end-to-end architecture for realtime ring VTO designed for deployment on 5G edge devices. The proposed solution leverages edge computing to offload computational tasks from user equipment, thereby reducing latency and improving the real-time VTO result fluidity through higher frame rates. We address the primary challenge of optimizing the 3D hand reconstruction block, which is the most timeconsuming component in the VTO pipeline. We explore various deep learning optimization techniques, including quantization, compilation optimization, and inference enhancement, to improve model performance on edge devices. Experimental results demonstrate significant improvements in inference speed, GPU memory usage, and storage efficiency. These enhancements validate the effectiveness of the proposed optimizations compared to the baseline setup. This work paves the way for more efficient and immersive VTO applications in 5G-enabled edge environments, enhancing the online shopping experience and increasing sales conversion rates.