Web-based DNNs provide accurate object recognition to the mobile Web AR, which is newly emerging as a lightweight mobile AR solution. Webbased DNNs are attracting a great deal of attention. However, balancing the UX against the computing cost for DNN-based object recognition on the Web is difficult for both self-contained and cloud-based offloading approaches, as it is a latency-sensitive service but also has high requirements in terms of computing and networking abilities. Fortunately, the emerging 5G networks promise not only bandwidth and latency improvement but also the pervasive deployment of edge servers which are closer to the users. In this article, we propose the first edge-based collaborative object recognition solution for mobile Web AR in the 5G era. First, we explore the finegrained and adaptive DNN partitioning for the collaboration between the cloud, the edge, and the mobile Web browser. Second, we propose a differentiated DNN computation scheduling approach specially designed for the edge platform. On one hand, performing part of DNN computations on mobile Web without decreasing the UX (i.e., keep response latency below a specific threshold) will effectively reduce the computing cost of the cloud system; on the other hand, performing the remaining DNN computations on the cloud (including remote and edge cloud) will also improve the inference latency and thus UX when compared to the self-contained solution. Obviously, our collaborative solution will balance the interests of both users and service providers. Experiments have been conducted in an actually deployed 5G trial network, and the results show the superiority of our proposed collaborative solution.

中文翻译：

---

用于5G网络中移动Web增强现实的边缘辅助分布式DNN协作计算方法

基于Web的DNN为移动Web AR提供准确的对象识别，这是一种新兴的轻量级移动AR解决方案。基于Web的DNN吸引了大量关注。但是，对于自包含的和基于云的卸载方法，要平衡UX与Web上基于DNN的对象识别的计算成本都是困难的，因为它是对延迟敏感的服务，但在计算和计算方面也有很高的要求。联网能力。幸运的是，新兴的5G网络不仅有望提高带宽和延迟，而且还可以广泛部署更靠近用户的边缘服务器。在本文中，我们提出了5G时代第一个针对移动Web AR的基于边缘的协作对象识别解决方案。第一，我们探索了用于云，边缘和移动Web浏览器之间协作的细粒度和自适应DNN分区。其次，我们提出了一种专门为边缘平台设计的差分DNN计算调度方法。一方面，在不降低UX（即将响应等待时间保持在特定阈值以下）的情况下，在移动Web上执行DNN计算的一部分将有效降低云系统的计算成本；另一方面，与独立解决方案相比，在云（包括远程云和边缘云）上执行其余的DNN计算也将改善推理延迟，从而改善UX。显然，我们的协作解决方案将平衡用户和服务提供商的利益。实验是在实际部署的5G试用网络中进行的，