Reproduce “Segment Anything Model for Road Network Graph Extraction” 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
Congrui Hetang, Haoru Xue, Cindy Le, Tianwei Yue, Wenping Wang, Yihui He
SAM-Road is a two-stage method leveraging a fine-tuned Segment Anything Model to extract vectorized road network graphs from satellite imagery. This approach achieves state-of-the-art APLS scores on City-scale (68.37%) and SpaceNet (71.64%) datasets, while being up to 40 times faster than previous methods due to its parallelized sliding-window inference.
Key insights
- Foundational vision models like SAM can be effectively adapted and fine-tuned to complex, specialized graph-learning tasks, offering robust visual understanding for structured prediction.
- Decoupling geometry prediction (via segmentation) from topology inference (via a learned GNN) effectively addresses the limitations of heuristic-based post-processing in traditional methods.
- A straightforward, minimalist design built upon a pre-trained foundational model can achieve state-of-the-art accuracy with exceptional computational efficiency and scalability for large-scale mapping.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.