CenterlineNet: Patch-Aligned Supervision For Thin Road Centerline Extraction

ICLR 2026 Conference Submission20149 Authors

Keywords: Road centerline extraction, Remote sensing imagery, Spatial misalignment, Patch alignment loss, Softmax-in-group weighting, Intersection-aware loss, Topology preservation, Weakly supervised segmentation, Vector field correspondence, CenterlineNet

TL;DR: CenterlineNet addresses spatial misalignment in road extraction by integrating patch alignment, reciprocal weighting, and intersection-aware loss to preserve connectivity.

Abstract:

Road networks evolve over time, requiring frequent map updates. AI tools can
assist with this task; however, methods based on raster segmentation followed by
thinning, skeletonization, or automatic tracing may fail to capture the local struc-
ture of road networks, increasing the burden on human annotators. Our goal is to
directly predict thin centerline representations that reflect structural patterns used
by annotators, particularly at intersections. A secondary goal is to scale train-
ing by learning from variable-quality vector data, such as OpenStreetMap, rather
than relying on precisely aligned segmentation masks that are difficult to produce
at scale. A key challenge is spatial misalignment in training data: while minor
for thick segmentation masks, even small shifts become a major obstacle when
learning thin centerlines, as pixel-wise losses are disproportionately affected. We
propose CenterlineNet, a weakly supervised model that addresses this challenge
with a patch alignment loss that compares local neighborhoods instead of individ-
ual pixels. This loss matches each predicted neighborhood to its nearest annotated
centerline, enabling flexible alignment within a distance tolerance. We present
two variants, basic and reciprocal, with the latter handling many-to-one mappings
via softmax-in-group weighting, and introduce an intersection-aware component
that specifically targets road junctions to improve connectivity.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 20149