Research map / Point clouds / Point cloud classification

CitySTAR: Agent-Driven Structured and Topology-Aware Reasoning for Open-Vocabulary Urban 3D Grounding

NeurIPS 2026 · Point clouds › Point cloud classification · approach: Point-language and open-vocabulary 3D

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.

Proposes CitySTAR, a training-free graph-based chain-of-thought framework for open-vocabulary city-scale 3D grounding that combines CodeLLM tool orchestration, paired-hypergraph topology verification, and 2D visual reranking, and introduces the CitySTAR-3D benchmark.

Problem

City-scale 3D visual grounding over billion-point urban point clouds with repeated objects and open-vocabulary targets, where spatial relations are hidden in global embeddings and matching-based methods fail.

Approach

Reformulates city-scale 3D grounding as graph-based chain-of-thought reasoning: a query-ready scene graph and CodeLLM-planned tool calls produce a high-recall candidate pool, paired query/candidate hypergraphs verify bidirectional target-context topology, and a candidate-centered 2D rendering is reranked by a visual-text model before the final selection.

Key results

• On CityRefer-NO, CitySTAR reaches 72.54 Acc@0.25 and 71.68 Acc@0.50, versus 50.69 and 46.86 for CityAnchor.
• On CityRefer-ND, CitySTAR reaches 82.67 Acc@0.25 and 80.68 Acc@0.50.
• On CitySTAR-3D, macro-average category Acc@0.25 is 56.88% versus 23.38% (CityRefer) and 22.41% (CityAnchor).

Surprising findings

• Swapping the CodeLLM backbone (Qwen3-Coder vs Qwen3.5-9B) changes Acc@0.50 only marginally, suggesting gains do not stem mainly from the LLM.

Limitations

• Depends on quality of open-vocabulary instance extraction in cluttered scenes.
• Limited spatial relation set; finer metric, functional, and commonsense relations not covered.
• CodeLLM tool use and visual reranking add computational overhead.

Experimental setup

Datasets: CityRefer CityAnchor SensatUrban STPLS3D

Benchmarks: CityRefer CityAnchor CitySTAR-3D

Metrics: Acc@0.25 Acc@0.50

Baselines: InstanceRefer 3DVG-Transformer EDA CityRefer CityAnchor

Builds on: CityRefer CityAnchor SAM3

Closest papers by meaning

• Reasoning in Space via Grounding in the World (ICLR 2026 · similarity 0.90)
• S$^2$-MLLM: Boosting Spatial Reasoning Capability of MLLMs for 3D Visual Grounding with Structural Guidance (CVPR 2026 · similarity 0.90)
• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.89)
• Grounded 3D-Aware Spatial Vision-Language Modeling (CVPR 2026 · similarity 0.88)
• G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding (CVPR 2024 · similarity 0.88)
• From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs (CVPR 2026 · similarity 0.88)