DeLuca Industries

ARGUS: Aerial Review & Ground-imagery Understanding for Search Argus is the second set of eyes for missing-person search teams, built as an open workflow rather than a single-image analyzer. Every day, 2,300 people go missing in America. A single drone search can generate thousands of photos, and tired volunteers review them by hand, often after dark. Review is the bottleneck of the search. A person can sit visible in imagery nobody had the hours to study. ARGUS exists to close that gap. It reviews a search team's own drone and grid imagery, ranks the candidates that could be the missing person, and pins them on the search map for a human to verify, so a reviewer checks ranked candidates instead of thousands of raw photos. The workflow runs end to end in a local web UI. First, ingest: a search team loads drone photos or video from a search grid. EXIF GPS is extracted and photos are placed on the map, video is split into one frame per second, and each frame is tiled for detailed review. Next, scan: SAM 3.1 segments each tiled frame and returns boxes, masks, and confidence scores. Then, verify: Muse Spark checks each top candidate crop, cropped from the original photo rather than an upscaled tile, against a written case brief, clothing, build, and last-known description and returns a 0–1 score with a one-sentence reason. Finally, review: ranked pins appear on the map and ranked cards appear in the Review list, where a human accepts or rejects each candidate and accepted pins export as CSV and GeoJSON for the ground team's map. A small Review Agent sits on top of that workflow and proposes, never decides. It dedupes overlapping-tile duplicates, re-checks borderline Muse scores from 0.4 to 0.6 with a second pass while keeping the original score, and writes a plain-text summary citing tile, crop, SAM confidence, and Muse score. A benchmark harness is also included and runs offline against a hand-written truth file containing only photos with known ground truth; a missed photo is never dropped to improve a score. ARGUS is built with SAM 3.1 for segmentation and Muse Spark for vision verification via the Meta Model API. Users can run it in about five minutes from the public repository: clone it, run setup, paste their own Meta key in the UI key field, start the UI, click USE DEMO PHOTOS, type a brief such as "adult person, dark shirt," and click Run. No terminal use is needed after setup, and bundled sample photos mean no search imagery is required to test it. The demo version can easily be removed and have ARGUS ready for deployment quickly. Enhancements move horizontally and can be implemented easily.