Research Presentation at EuCNC & 6G Summit
M2 Student Maruyama has presented his research at EuCNC & 6G Summit
Publications & Talks
International Conference · 2025
Design and Implementation of Point Cloud-Based Space Integration with Quality-Aware Optimization
Abstract
We present a hybrid-metaverse system that integrates physical and virtual spaces through real-time point cloud processing. To address the challenges of high-bandwidth data transmission and quality management across multiple users, we introduce a ROS-based distributed architecture that combines Input Convex Neural Networks with ADMM optimization. Evaluation using the SUN RGB-D dataset under simulated 6 G conditions demonstrates our system's effectiveness, with PICNN achieving 93-94.6% accuracy in modeling user utility and the distributed approach reducing convergence time by up to 60% compared to centralized methods while maintaining sub-millisecond latency with 30 concurrent users. The results validate our framework's capability to manage real-time point cloud quality in next-generation network environments.
Research Note
A hybrid metaverse that captures real spaces as point clouds and merges them with the spaces of distant participants runs into a bandwidth problem. Denser point clouds look better and cost proportionally more to transmit, so once several users share a link they are competing for it, and someone has to decide how much quality each of them receives.
Satisfaction does not rise linearly with the bandwidth allocated. It improves quickly at first and then flattens. A utility function expresses that relationship between allocation and satisfaction. An Input Convex Neural Network is a network constrained by construction so that its output is convex in its input, which allows the utility function to be learned from data while keeping the allocation problem that follows solvable as a convex program.
This work combines such an ICNN with ADMM optimization on a ROS-based distributed architecture. ADMM splits a large optimization problem into subproblems that are solved in parallel and reconciled into a global solution, which responds faster than centralising the whole computation.
Evaluated on the SUN RGB-D dataset under simulated 6G conditions, the ICNN modelled user utility with 93 to 94.6% accuracy, and the distributed approach cut convergence time by up to 60% relative to centralised methods while holding latency below a millisecond with 30 concurrent users.
Related Lab Activities
