Abstract the Layout, Focus the Detail: A Dual-Granularity Representation Framework for Zero-Shot 3D Visual Grounding

Official repository for: Abstract the Layout, Focus the Detail: A Dual-Granularity Representation Framework for Zero-Shot 3D Visual Grounding, Accepted at ECCV 2026.

Overview

We present a dual-granularity framework for zero-shot 3D visual grounding. A Semantic Spatial Layout abstracts global scene structure for spatial reasoning, while Visual Detail Patches preserve local appearance cues for visual verification. Together with query-guided scene filtering, the two forms of evidence support joint spatial-visual reasoning for target grounding.

Citation

@inproceedings{lin2026abstract,
  title={Abstract the Layout, Focus the Detail: A Dual-Granularity Representation Framework for Zero-Shot 3D Visual Grounding},
  author={Lin, Zeyuan and Li, Hanxuan and He, Chen and Wang, Ruiping and Liu, Zhaoxiang and Chen, Xilin},
  booktitle={European Conference on Computer Vision},
  year={2026}
}