Reproduce “PointScatter: Point Set Representation for Tubular Structure 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
Dong Wang, Zhao Zhang, Ziwei Zhao, Yuhang Liu, Yihong Chen, Liwei Wang
PointScatter replaces fixed-grid masks with locally predicted point sets for extracting thin, branching tubular structures. Across retinal vessel, road, and satellite datasets, it improves volumetric accuracy and especially centerline recall, while auxiliary point-set supervision further strengthens conventional segmentation models; connectivity errors remain a limitation.
Key insights
- A point representation gives the model greater flexibility for recovering fine-scale structures, with the strongest gains observed for one-pixel-wide centerlines and small branches.
- PointScatter improves volumetric metrics and recall, but recovering additional thin segments can introduce disconnected components and therefore does not consistently improve topology-based measures.
- Jointly training a point-set branch with mask prediction can transfer useful geometric supervision to the shared backbone, improving both segmentation quality and selected topological metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.