Reproduce “DenseGrounding: Improving Dense Language-Vision Semantics for Ego-Centric 3D Visual Grounding” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Henry Zheng, Hao Shi, Qihang Peng, Yong Xien Chng, Rui Huang, Yepeng Weng, Zhongchao Shi, Gao Huang

Tsinghua University and Lenovo Research developed DenseGrounding, an approach for ego-centric 3D visual grounding that addresses the loss of fine-grained visual details and ambiguous language descriptions. The method achieves state-of-the-art performance on the EmbodiedScan benchmark, improving overall accuracy by 5.81%.

Key insights

  • Preserving and enhancing visual semantics at multiple levels (view-level to point cloud infusion) is crucial for accurate 3D visual grounding, especially for small objects.
  • Enriching ambiguous natural language descriptions with LLM-generated contextual anchors and scene information significantly improves the ability to disambiguate between similar objects.
  • The synergistic combination of enhanced visual and textual semantics leads to robust performance gains across various object categories and scene types.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.