MM-Conv: A Multimodal Dataset and Benchmark for Context-Aware Grounding in 3D Dialogue

2026-05-20

AI Summary

Existing vision-language models struggle with ambiguous referring expressions in dynamic, multi-turn dialogues, particularly lacking context-aware referential grounding capabilities in 3D environments. This work proposes the first multimodal dialogue referring benchmark tailored for dynamic 3D scenes, comprising 6.7 hours of first-person VR interaction data that integrates speech, actions, gaze, and 3D geometric information. The authors introduce a two-stage approach: first disambiguating linguistic references through context-aware utterance rewriting, then performing visual grounding using models such as GroundingDINO. Experimental results demonstrate that this decoupled strategy substantially outperforms end-to-end baselines, with context rewriting yielding consistent performance gains of 11–22 percentage points. Notably, the method achieves a 56.7% accuracy on pronoun-based reference tasks—nearly double that of the strongest baseline.

Abstract

Grounding language in the physical world requires AI systems to interpret references that emerge dynamically during conversation. While current vision-language models (VLMs) excel at static image tasks, they struggle to resolve ambiguous expressions in spontaneous, multi-turn dialogue. We address this gap by introducing (1) a benchmark for referential communication in dynamic 3D environments, built from 6.7 hours of egocentric VR interaction with synchronized speech, motion, gaze, and 3D scene geometry, and (2) a two-stage grounding pipeline that explicitly resolves conversational ambiguity before visual localization. The benchmark includes over 4,200 manually verified referring expressions spanning full, partitive, and pronominal types. Our contextual rewriting approach improves grounding performance by 11-22 percentage points on average, with a pure detector (GroundingDINO) reaching 56.7% on pronominals after rewriting, nearly double the best end-to-end baseline. Results demonstrate that decoupling linguistic reasoning from visual perception is more effective than end-to-end approaches for conversational grounding.

Problem

• conversational grounding
• referring expressions
• multimodal dialogue
• 3D environments
• context-awareness

Innovation

• multimodal grounding
• context-aware dialogue
• 3D referential communication
• conversational ambiguity resolution
• two-stage grounding pipeline