Rust-accelerated core
Tile stitching, polygon rasterization, and patch sampling all run in compiled Rust via PyO3, which is orders of magnitude faster than other implementations.
mapcv turns a bounding box and a set of polygon labels into a ready-to-train image segmentation dataset. It fetches XYZ map tiles, rasterizes KML or GeoJSON annotations onto the tile grid, extracts fixed-size image/mask patches, and writes them to disk. To ensure a lightweight footprint and easy installation, the tool is GDAL-free. Tile stitching, rasterization, and patch sampling are powered by Rust, accessible through the CLI (or via the Python API).
Many tile sources
Built-in support for Esri World Imagery, Google Satellite, OpenStreetMap, CartoDB basemaps, and any custom XYZ URL template.
KML & GeoJSON labels
Parse polygon annotations directly from KML or GeoJSON files. Multiclass labels via a configurable field name. Automatic WGS-84 to Web Mercator projection.
Flexible patch sampling
Grid or random sampling with configurable stride, edge strategies (pad / drop / shift), and empty-patch filtering.
Train / val / test splits
Stratified or random dataset splitting from the JSON manifest. Configurable labeled-data fractions for semi-supervised workflows.
Simple CLI
Run the full pipeline with mapcv generate config.yaml. One command, one config file, ready-to-train output.