16 Paper Presentations at ICDCN2026
16 presentations and 2 awards at ICDCN 2026 held in Nara, Japan (January 2026)
Publications & Talks Awards
International Conference · 2026
When Sensors Talk: LLM-Guided Object-Centric Collaboration for Distributed 3D Scene Understanding
Abstract
Collaborative 3D scene understanding enables multiple LiDAR or depth sensors to jointly perceive and reason about shared spaces, but transmitting full point clouds across devices is impractical under real-world bandwidth and latency constraints. Existing large language model (LLM)–based frameworks also typically assume access to a single, holistic 3D view. We propose an object-centric collaborative reasoning framework in which each device encodes local objects, shares lightweight summaries for multi-view fusion, and performs high-level reasoning through an LLM interface. Experiments on dense indoor scenes show that our method achieves accuracy close to raw-data fusion while greatly reducing communication bandwidth.
Research Note
A single LiDAR or depth sensor cannot see what is hidden behind an object. Placing several sensors so that they cover each other's blind spots solves the geometry, but sharing raw point clouds between them does not fit within realistic bandwidth and latency budgets, since a point cloud amounts to hundreds of thousands of points per second.
The useful observation is that what needs to be shared is not the points but the meaning behind them. An object-centric representation describes a space as a set of objects, one desk and two chairs and a person, rather than as a cloud of measurements. It discards most of the data while keeping what later reasoning actually depends on.
In the proposed framework each device encodes the objects it can see locally and shares only a lightweight summary. Summaries arriving from several viewpoints are fused into a single account of the scene, and higher-level reasoning over that account is carried out through a large language model interface. This differs from earlier LLM-based frameworks, which generally assume access to one holistic 3D view, by taking distributed viewpoints as the starting point.
Experiments on dense indoor scenes achieved accuracy close to fusing the raw data while greatly reducing the communication required.
Related Lab Activities
Related Projects
セマンティック通信による多端末連携型の状況理解と消防システムへの適用
Ministry of Internal Affairs and Communications (MIC) FORWARD
デジタルインフラ構築部門
