Autobounds: Pixels to Polygons

Build a web map that fetches imagery, runs a segmentation model on Cloud Run, converts masks to GeoJSON, calculates areas, and stores results with Firebase.

Overview

Autobounds is a geospatial web app that makes farm maps actionable by combining real‑time NDVI visualization with fast, AI‑powered field boundary extraction. This demo‑driven talk walks through the full stack:

Frontend: MapLibre GL layers for NDVI, field labels, and UX for extract/review/save with geodesic area, undo/redo, and editing.
Backend: a “Delineate Anything” segmentation service containerized in Python/Flask, deployed to Google Cloud Run; Firebase Functions act as a backend‑for‑frontend for auth, validation, and CORS.
Data flow: capture map bounds → fetch imagery → infer polygons → convert masks to GeoJSON → compute areas → persist to Firestore.
Hard‑won lessons: emulator vs production pitfalls, OpenCV/libGL in Docker, $PORT expansion on Cloud Run, CORS, deterministic inference, model loading and memory, and keeping the map responsive at various zooms.

Attendees will leave with practical patterns to ship geospatial ML features to production:

How to stream model results into browser map layers with good UX.
How to package and run heavy models on Cloud Run with minimal cold‑start pain.
How to use Firebase Functions as a thin BFF to shield the frontend and simplify auth/CORS.
How to test E2E locally with reproducible debug bundles.

Links

Tech stack