Enterprise AI Analysis

Current Scalability Challenges in 3DSGG

Current 3D scene graph generation (3DSGG) approaches are fundamentally limited by their reliance on single-agent paradigms and their evaluation within small-scale, constrained environments. This leads to significant scalability challenges when applied to real-world, large-scale scenarios. For instance, runtimes can be up to 4x longer and data traffic 98x heavier compared to multi-agent solutions. The absence of comprehensive benchmarks for multi-agent, large-scale, and dynamic environments further exacerbates these limitations, hindering the development of truly scalable AI for complex domains.

Decentralized Multi-Agent 3DSGG Framework

MA3DSG introduces a novel decentralized multi-agent framework designed for scalable and efficient 3D scene graph generation in large-scale environments. It comprises three core components:

• Multi-Agent Exploration: Enables distributed and collaborative exploration of diverse regions.
• 3D Semantic Scene Graph Generation: Agents incrementally build local 3D semantic scene graphs from RGB-D sequences, utilizing a 3D Global Segmentation Map (GSM) and Feature Graph (SGFN-based).
• 3D Semantic Scene Graph Alignment: A lightweight, training-free algorithm that efficiently merges partial query graphs from individual agents into a unified global scene graph, enhancing completeness and reducing overhead through overlapping exploration and robust node/edge attribute updates.

This design allows conventional single-agent systems to operate collaboratively, ensuring fast and robust performance without requiring complex learnable parameters for alignment.

A Comprehensive Benchmark for Scalable 3DSGG

To rigorously evaluate 3DSGG performance and scalability, MA3DSG-Bench provides a flexible and extensible benchmark with diverse agent configurations (single/multi-agent), varying scales (1 to 47 rooms), and scene dynamics (static and long-term changes). Unlike previous benchmarks that process each room independently, MA3DSG-Bench treats all reference rooms as a unified navigable space, enabling parallel exploration. It also incorporates rescan sequences to reflect realistic long-term environmental changes, facilitating joint perception and temporal context modeling in dynamic multi-agent environments. This benchmark sets a new standard for evaluating scalable 3DSGG in real-world conditions.

Superior Scalability and Robustness

Our experiments confirm MA3DSG's superior scalability and efficiency. In Static Collaborative Perception (SCP) scenarios, MA3DSG achieves comparable performance to single-agent baselines while running up to 4x faster and using 98x less data traffic than multi-agent baselines in large-scale environments. In Long-term Dynamic Collaborative Perception (LDCP) scenarios, MA3DSG adeptly manages temporal inconsistencies and updates scene graphs with comparable F1@1 scores for triplets, objects, and predicates, while operating 3.4x to 4.1x faster. Its lightweight graph alignment algorithm ensures consistently low latency and significantly reduces communication overhead, making it ideal for real-world, large-scale deployments.