SEMIR: Topology-Preserving Graph Minors for Thin-Structure Segmentation
SEMIR replaces pixel-level segmentation with classification over a compact graph minor. It collapses millions of pixels into hundreds of boundary-aligned supernodes while preserving the connectivity of thin structures like power lines or cracks. This allows for high-resolution inference using a lightweight Graph Neural Network, achieving state-of-the-art results on multiple thin-structure datasets.