SEMIR: New graph method segments thin structures 4.6x cleaner

Thin-structure segmentation—power lines, pavement cracks, or lane markings at just 1–3 pixel width—has long been a thorn in computer vision. Standard approaches either patch the image (breaking continuous lines) or use superpixels that often merge thin targets into background before classification. Topology-aware losses can penalize connectivity breaks, but they cannot recover what the representation has already destroyed. Enter SEMIR, a framework from researchers Luke James Miller and Yugyung Lee (accepted at ECCV 2026). SEMIR sidesteps the pixel lattice entirely by building a parameterized graph minor whose contraction map preserves thin-structure connectivity. It collapses millions of pixels into tens or hundreds of boundary-aligned supernodes, enabling full-resolution inference without patching—demonstrated at scales up to 21 megapixels. A lightweight graph neural network (GNN) then classifies the reduced graph, and an exact map lifts predictions back to pixel resolution.

The results are striking: using a single pipeline with identical architecture, features, loss, and GNN hyperparameters across all datasets, SEMIR matches or exceeds several domain-specific baselines on the TTPLA (power lines), CrackSeg9k (pavement cracks), and SkyScapes Lane (aerial markings) benchmarks. It achieves competitive Dice, IoU, and Boundary F1 scores while reducing mask fragmentation by at least 4.6× relative to the widely-used SLIC superpixel method. This means models can now process high-resolution imagery in one pass, preserving topological integrity without expensive post-processing. For applications like autonomous driving (lane detection), infrastructure inspection (crack monitoring), and utility mapping (power line tracking), SEMIR offers a practical path to more reliable, less fragmented segmentation.