Reproduce “Vector Map as Language: Toward Unified Remote Sensing Vector Mapping” 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
Yinglong Yan, Yunkai Yang, Haoyi Wang, Wei Fu, Linshan Wu, Honghu Pan, Shaobo Xia, Shanghang Zhang and 2 more
Researchers from Hunan University and collaborators propose VecLang, a unified paradigm that redefines multiclass remote sensing vector mapping as structured text generation via a GeoJSON-like Structured Vector Language. The framework outperforms specialized methods in multiclass scenarios and shows robust generalization across datasets and to unseen object categories, achieving a mean score of 78.22 for open-vocabulary tasks.
Key insights
- Reformulating multiclass vector mapping as structured text generation via a GeoJSON-like language enables a single model to unify the description of diverse geospatial entities, including both closed and network-like structures.
- A progressive framework that localizes vectorization units and generates structured text incrementally significantly improves efficiency, parse rates, and overall map accuracy in complex remote sensing scenes.
- Hierarchical reinforcement learning, incorporating syntax, content, and executability rewards, is essential for guiding generative models to produce geometrically precise, semantically accurate, and topologically consistent vector maps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.