[Paper Review] Immersion on the Edge: A Cooperative Framework for Mobile Immersive Computing

Zeqi Lai, Yong Cui+2 more|arXiv (Cornell University)|Jul 12, 2018

Computer Graphics and Visualization Techniques6 references4 citations

TL;DR

This paper proposes CoIC, a cooperative framework that reduces latency in mobile immersive computing by caching and sharing intermediate results—such as DNN features and 3D model hashes—across users and applications at the mobile edge. By leveraging computation redundancy in VR/AR workloads, CoIC achieves up to 52.28% reduction in object recognition latency and 75.86% in 3D rendering load latency under real-world network conditions.

ABSTRACT

Immersive computing (IC) technologies such as virtual reality and augmented reality are gaining tremendous popularity. In this poster, we present CoIC, a Cooperative framework for mobile Immersive Computing. The design of CoIC is based on a key insight that IC tasks among different applications or users might be similar or redundant. CoIC enhances the performance of mobile IC applications by caching and sharing computation-intensive IC results on the edge. Our preliminary evaluation results on an AR application show that CoIC can reduce the recognition and rendering latency by up to 52.28% and 75.86% respectively on current mobile devices.

Motivation & Objective

• Address the growing demand for low-latency, high-quality immersive computing (IC) applications like VR and AR on mobile devices.
• Identify and exploit computation redundancy in IC workloads across users and applications in close proximity.
• Improve end-to-end user quality of experience (QoE) by reducing latency through cooperative caching at the mobile edge.
• Design a lightweight, scalable framework that enables real-time reuse of intermediate IC results without compromising accuracy.

Proposed method

• Use feature descriptors—such as DNN-embedded feature vectors or hash values of 3D models/panoramas—as keys to index and match intermediate IC results in edge caches.
• Implement a two-tier request flow: first check the edge cache using the feature descriptor; if miss, forward to the cloud and cache the result for future reuse.
• Apply similarity-based matching (e.g., L2 distance thresholding) to determine if a new input matches a cached result, even with minor variations.
• Deploy the client on a Pixel smartphone with Android O, and run edge and cloud modules on Linux machines, using 802.11ac WiFi for client-edge communication.
• Use tc to emulate real-world mobile network conditions, enabling controlled evaluation of latency under varying bandwidth.
• Focus on caching for recognition and rendering tasks, excluding object tracking due to its lower computational cost and feasibility on mobile devices.

Experimental results

Research questions

• RQ1 To what extent can computation redundancy across users and applications in mobile immersive computing be exploited to reduce latency?
• RQ2 How effective is edge caching of intermediate IC results (e.g., DNN features, 3D model data) in improving end-to-end latency?
• RQ3 What is the impact of network variability on the performance gains of cooperative caching in mobile IC systems?
• RQ4 How can a lightweight, scalable framework be designed to enable real-time reuse of IC results without introducing significant overhead?

Key findings

• CoIC reduces object recognition latency by up to 52.28% by reusing cached DNN feature vectors when similar inputs are detected across users.
• For 3D model rendering, CoIC achieves up to 75.86% reduction in load latency by caching preloaded model data on the edge.
• The framework demonstrates significant performance gains under diverse network conditions, with latency reduction sustained even at lower bandwidths.
• The use of feature descriptor matching (e.g., L2 distance thresholding) enables accurate and efficient reuse of results despite input variations.
• The system shows strong potential for real-world deployment, as it operates within realistic mobile and edge network constraints.
• The results validate the hypothesis that computation redundancy in mobile IC workloads is both prevalent and exploitable for QoE improvement.