bag-lab/MAP-3M

Dataset Name

MAP-3M - A Large Scale Multi-Class Map Dataset

Dataset Summary

MAP-3M is one of the largest high-resolution aerial imagery and map datasets to date, comprising approximately 3 million images—10× larger than comparable datasets. Each image is enriched with high-quality annotations for two fundamental map classes: buildings and roads.

Images: Sourced from the National Agriculture Imagery Program (NAIP) (U.S. Department of Agriculture, 2025). Sampling: Leveraging population data from the United States Cities Database (2025), we evenly sample 5,000 cities across all 50 states. Labels: Vectorized annotations provided in COCO format, covering buildings and roads.

Supported Tasks and Leaderboards

Tasks:

Map Generation Semantic Segmentation Classification

Leaderboards: TBD – ICLR 2026

Dataset Structure

We provide the annotation in COCO style dataset.

Train

1. coco_train_interpolated_60_filtered.json
2. coco_train_interpolated_60_filtered.ndjson

Val

1. coco_val_interpolated_60_filtered.json
2. coco_val_interpolated_60_filtered.ndjson

Instructions

zip -s 0 MAP3M.zip --out MAP3M_full.zip unzip MAP3M_full.zip

Citation

@dataset{MAP-3M, author = {Anonymous}, title = {MAP-3M: Large Multi-Class Map Dataset}, year = {2025}, url = {https://huggingface.co/datasets/bag-lab/MAP-3M} }

Acknowledgements

We thank the U.S. Department of Agriculture for NAIP imagery and the United States Cities Database for population data. Special thanks to all contributors for dataset preparation and annotation.