Model ICDAR 2021 24M

MapSeg21 Baseline U-Net for Historical Map Segmentation

Reported evaluation

Dataset Split Metric Score
MapSeg21 test COCO PQ (building blocks) 0.62
MapSeg21 test mean IoU (map content) 0.91

Downloads

File Type Size Mirrors
mapseg-baseline-unet.pt
Pre-trained PyTorch weights (baseline reference)
weights 92 MB Mirror 1
README.md
Model card and usage instructions
model-card Mirror 1

Description

A reference U-Net baseline released alongside the ICDAR 2021 MapSeg competition. The model is trained to perform two of the three competition tasks (building-block detection and map-content segmentation) on 1894-1937 atlas sheets of the City of Paris.

It is published primarily as a sanity-check baseline so that participants in any future re-run of the competition can confirm their data loading and evaluation pipeline against a known reference, not as a state-of-the-art model.

See the companion dataset and the evaluation tools for the full reproducible pipeline.

Note: this entry is illustrative for the catalog prototype; the weights file listed above is a placeholder URL, not a real download.