ArXiv paper probes thin-structure boundary metrics

Smart focus on a real failure mode: average overlap scores can look fine while thin structures break in practice.

• –The paper’s core contribution is an evaluation protocol, not a new model, which is useful for reproducible benchmarking.
• –Comparing BF1 and Boundary-IoU against F1/IoU helps surface contour-quality tradeoffs that standard metrics can hide.
• –The reported classical-vs-learned tradeoff (higher mean vs better worst-case reliability) is practical for deployment decisions.
• –Small-scale setup (single task, limited held-out set) makes this more of a strong methodology note than a definitive benchmark.