Research map / 3D vision / 3D generation

CAIRN: Cross-Room 3D Scene Understanding with Topology-Aware Large Multimodal Models

NeurIPS 2026 · 3D vision › 3D generation · approach: Open-vocabulary 3D scene understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CAIRN, a topology-aware 3D-LLM using hierarchical scene graph tokens, room tokens, hierarchical masked attention and geometric bias, plus the CAIRN-MR multi-room benchmark built on HM3D.

Problem

Existing 3D-LLMs are designed for single-room scenes and represent scenes as flat object-token sequences, which fails to reason over multi-room household environments with cross-room dependencies.

Approach

Encode a multi-room scene as a two-layer hierarchical scene graph, add GNN-enriched object tokens and learned room tokens, and apply an attention mask that routes information by room topology plus a learned geometric bias on attention logits.

Key results

• On CAIRN-MR with Qwen3-8B, CAIRN reaches 25.8 Acc@0.25 grounding and 94.5 CIDEr captioning, versus 20.4 and 79.5 for the best baseline Inst3D-LMM.
• CAIRN gains +6.7 EM on inter-room comparison and +3.4 EM on existence over the strongest baseline on CAIRN-MR.
• Ablation: removing the hierarchical mask drops comparison EM from 70.4 to 64.3 and captioning CIDEr from 94.5 to 85.7.

Surprising findings

• Counting frequency baseline reaches 62.6 MRA, close to some trained models, due to skew toward small integers in indoor scenes.
• Identification accuracy is essentially unchanged across methods (about 62 EM), indicating room-level structure helps less for within-room object recognition.

Limitations

• Scale of model capacity and training data is limited; authors note further scaling may help.
• CAIRN-MR covers only 673 scenes from HM3D and does not extend to more complex multi-level environments.
• Reasoning is implicit; no explicit multi-step reasoning process is modeled.

Experimental setup

Datasets: HM3D SceneVerse ScanNet ScanRefer Multi3DRefer Scan2Cap ScanQA SQA3D

Benchmarks: CAIRN-MR ScanRefer Multi3DRefer Scan2Cap ScanQA SQA3D

Metrics: Acc@0.25 Acc@0.5 F1@0.25 F1@0.5 CIDEr BLEU-4 exact match MRA

Baselines: Chat-Scene 3DGraphLLM Inst3D-LMM GPT4Scene-HD LEO Scene-LLM Grounded 3D-LLM PQ3D

Builds on: Chat-Scene Mask3D SceneVerse Graphormer ALiBi

Closest papers by meaning

• LSceneLLM: Enhancing Large 3D Scene Understanding Using Adaptive Visual Preferences (CVPR 2025 · similarity 0.89)
• SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding (NeurIPS 2026 · similarity 0.88)
• Chat-Scene: Bridging 3D Scene and Large Language Models with Object Identifiers (NeurIPS 2024 · similarity 0.88)
• TB-HSU: Hierarchical 3D Scene Understanding with Contextual Affordances (AAAI 2025 · similarity 0.88)
• MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations (NeurIPS 2024 · similarity 0.88)
• Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning (CVPR 2025 · similarity 0.87)