コロキアムB発表

日時: 06月19日 (Thu) 3限目(13:30-15:00)


会場: L1

司会: 嶋利 一真
KANG XINGYUAN D, 中間発表 ソフトウェア設計学 飯田 元, 藤川 和利, 市川 昊平, 柏 祐太郎
titile: Enhancing Distributed Software Defined Networking: Multi-objective Controller Placement
abstract: In distributed Software Defined Networking (SDN), multiple controllers must maintain a consistent network state using consensus algorithms, which introduces additional communication overhead and delay, particularly in large-scale deployments. As a result, optimizing controller placement is challenging, as it requires consideration of both switch-to-controller latency and consensus-related delays. Furthermore, SDN controllers have limited processing capacities, and unbalanced workloads may result in high response times or even failures. To address these challenges, a multi-objective optimization method was designed to determine the optimal number and placement of controllers and datastore nodes. This study introduces two key metrics as the objectives for the optimization problem: Flow Setup Time (FST), a model measuring the overall network delay in distributed SDNs, and the Variance of Load Balancing (VOLB), which quantifies the degree of workload distribution across controllers. Evaluation on real network topologies from the Internet Topology Zoo demonstrated key trade-offs among the number of controllers, the number of datastore nodes, FST, and VOLB. Building on this foundation, ongoing research explores the application of machine learning-particularly reinforcement learning-in distributed SDN routing. By leveraging the logically centralized view of SDN, the research aims to improve routing performance through intelligent traffic engineering. This work targets ONOS-based architecture and investigates how learning-based methods can dynamically optimize routing decisions to minimize latency, balance load across controllers and datastores, and maximize network throughput.
language of the presentation: English
 
秋山 真哉 D, 中間発表 ユビキタスコンピューティングシステム 安本 慶一, 藤川 和利, 諏訪 博彦, 松田 裕貴
title: A Method for Optimizing Scanning Task Assignment for Accurate and Cost-Efficient 3D Model Generation via Photogrammetry
abstract: In recent years, the use of 3D models has been expanding in various fields such as virtual tourism, autonomous driving, and integrity checks of structures under construction. In these applications, it is essential to continuously and accurately update the 3D models of the target objects. Generally, methods using 3D LiDAR or 3D laser scanners can create highly accurate models, but the equipment is expensive and requires specialized knowledge to operate, making them unsuitable for frequent updates. On the other hand, photogrammetry, which creates 3D models from photographs, is a relatively accessible method that does not require specialized equipment. However, it faces challenges in securing a sufficient number and quality of photographs to ensure adequate accuracy. To address this issue, this study proposes a participatory photogrammetry approach that optimizes task requests to guide contributors in capturing photographs that are effective for creating high-fidelity 3D models. Our method begins by collecting a set of randomly sourced images (e.g., from the internet) to generate an initial coarse 3D model. Based on the estimated size and shape of this model, along with the specifications of typical smartphone cameras, the system calculates optimal viewpoints for model reconstruction. Individuals in the vicinity of the target object are then asked to take photos from these calculated positions, thus participating in the scanning process. Experiments show that 3D models generated from photos taken according to these optimized scanning tasks are not only visually more faithful to the real-world object compared to those created via standard photogrammetry, but also exhibit approximately 8.8% more vertices in the resulting point cloud, indicating an improvement in geometric detail and fidelity.
language of the presentation: Japanese