What I built
I developed a workflow for inspecting structure outlines, correcting mistakes, and preserving the decisions behind an exported map. Footprint Lab pairs a hands-on browser demonstration with supporting geospatial processing and review code.
A predicted outline can miss a roof, merge adjacent buildings, or mark something that is not a structure. I wanted a reviewer to correct those mistakes while keeping the original result and the history of each change available for inspection.
Try the review workflow
The browser demo uses an invented aerial-style scene and prepared synthetic outlines. Move a shifted boundary, replace a merged outline, add an omitted structure, or reject a false detection; then undo a change or download the GeoJSON and review history.
Edits stay in this browser’s local storage and can be recovered after a reload. The scene and coordinates are fictional, and no account or imagery upload is needed; the Python inference pipeline and React application remain separate source samples.
Architecture and decisions
- Keep corrections inspectable. Review decisions, geometry changes, and export metadata retain their history; the supporting application adds revision checks and idempotent operations for repeated or conflicting work.
- Process imagery in pieces. Raster windows, overlapping tiles, and batched inference address images larger than available memory, with geometry processing turning probability maps into individual outlines.
- Track reference-label provenance. Ancestry checks distinguish human reference labels from model-assisted corrections, keeping a corrected prediction from silently becoming independent reference evidence.
- Make review mistakes reproducible. A fixed synthetic scene gives visitors and tests the same shifted boundaries, merged outlines, omissions, and false detections.
My contribution
I initiated the project and directed its scope, design decisions, and reviews, using the AI-assisted development approach described on the About page.
Scope and verification
The reviewed revision passed demo and source-export CI; selected scientific-Python test results are retained project evidence, not a fresh inference benchmark. The sample demonstrates review behavior without establishing sensor accuracy, representative-site performance, operator adoption, or a deployed inference service.
Source snapshot 10580f2 · Reviewed 2026-09-22. This case study describes that revision; development may continue separately.