← All research

BYU FROST Lab · Aug 2025 – Present

What Transfers to Road Topology?

Vision foundation models are excellent at segmenting roads, but a road map also has to be connected. This project asks which pretrained representations actually carry over to extracting road networks as graphs from satellite imagery.

Alejandro Pelaez, Dallon Gunn, Brady Moon · First-author manuscript in preparation

Motivation

Road-network graphs extracted from satellite imagery feed autonomous navigation, urban planning, and disaster response. Getting the pixels right isn't enough: a single missed segment under a tree or overpass breaks every route that passes through it. Foundation models such as SAM, DINO, and RADIO have transformed pixel-level segmentation, but it was unclear which of their representational properties transfer to graph-level topology, where long-range connectivity matters as much as detection.

Approach: a controlled swap

To isolate the effect of the representation, the encoder is the only thing that changes. Everything downstream (decoder, topology reasoning, training recipe, thresholds) is held fixed.

Pipeline: satellite image to interchangeable encoder to feature map to decoder to road network graph
The unified framework. The encoder is swapped across experiments; the decoder and topology modules stay fixed.

Findings

69.3
APLS for SAM 2.1 on CityScale, the best of all six backbones (SAM: 65.8)
+4.3
Out-of-domain APLS gain for SAM 2.1 over SAM on unseen geographies
≈1.8×
Out-of-domain APLS of SAM 2.1 vs. the ResNet-50 baseline (40.4 vs. 22.4)
  1. The choice of backbone matters. SAM 2.1 gave the highest APLS, mainly by recovering roads the other encoders missed (higher recall).
  2. Foundation models generalize; supervised features don't. On unseen geographies the ResNet-50 baseline stays precise but its recall collapses, missing most of the network. Every foundation model held up far better.

The results are consistent with a simple hypothesis: pretraining objectives that emphasize spatial boundary localization (as in SAM 2.1) transfer better to topological routing than purely semantic or multi-teacher distilled objectives.

Grid of ground-truth road masks and predictions from SAM 2.1, SAM, DINOv3 and C-RADIOv3 on three scenes
Qualitative comparison on CityScale. Circles mark places where an encoder bridges (cyan) or breaks (red) an occluded connection.

My role

← Back to all research