chans/edgegate-package-entrance-eval

EdgeGate: package-at-the-entrance eval sessions

The evaluation recordings behind EdgeGate (Meta Global AI Developer Hackathon, Indie track, 2026-10-26 to 11-04): seven sessions, 2 h 50 min, 26 marked package arrivals, recorded with one OAK-D Pro fixed at an indoor drop zone that stands in for an entrance. Every build in the project was measured on its edge device over exactly these frames, and the gate held every one of them.

People are blurred. Every person is blurred (boxes from the project's own detector, widened by 12% and carried one second to each side; for u01 the uncovered frames were also checked for motion), and the monitor and the hallway are blurred in every frame (scripts/blur_sessions.py in the repository). Because the JPEGs were re-encoded, this public copy does not reproduce the eval set hash in the project's policy; manifests/ holds the per-frame sha256 of the originals, so a frame can be matched to its original by name, and the originals can be shared with reviewers on request.

What is here

sessions/<name>/frames/000000.jpg …   1280×720 JPEG frames, 15 fps, blurred as above
sessions/u01/frames/000/000000.jpg …  u01 only: 1,000 frames per folder, named by the first three
                                      digits (a folder on the Hub holds at most 10,000 files)
sessions/<name>/frames.csv            frame index and capture time in seconds (seq, t)
sessions/<name>/meta.json             lighting, note, camera, outages, duration
gt/<name>.json                        arrivals marked by hand: arrive, remove, box, tags, notes
manifests/<sha256>.json               the eval set manifests (v1 and v2) of the originals

session	lighting	length	marked arrivals	note
e01	night, one floor lamp (zone about 11/255)	4.8 min	2	cut short by a USB error; both boxes in the zone at the start, removed in the first 20 s
e02	night, one floor lamp	6.9 min	5
e03	night, room lights on	8.0 min	7 (+2 second packages, scored apart)	drops, walk-throughs, standing in the zone, boxes set below the zone
e04	night, room lights on	9.9 min	4
e05	rainy day, lights off	9.9 min	4
e06	night, room lights on	9.9 min	4	recorded after the first four students had been measured on e01–e05; fixed round script
u01	rainy day, unattended	119.9 min	0	no package placed; people walk through 44 times

Ground truth: arrive is when the hand lets go of the package, remove when it is picked up (null if it stays), box where it rests (normalized), tags such as edge (bottom-center within a few pixels of the zone line) or occluded. The notes field of each file records the placements that were removed as not arrivals under the zone rule (a box across the zone's bottom edge with its bottom-center outside), decided when the session was marked, before any candidate was measured on it.

The zone, the arrival rule and the thresholds are in the repository's policies/package_entrance.toml: a package whose box bottom-center is inside the zone, seen for 2 s in an empty zone, is an arrival; the alert must be received within 10 s of release.

Results on these sessions

Best student (s3, YOLOv8n distilled from SAM 3.1 labels on separate training recordings): recall 25/26 on both a Jetson Orin Nano (TensorRT FP16) and a Raspberry Pi 5 with a Hailo-8 (INT8), but 4.2 and 4.6 false alerts per hour against a limit of 1, so HOLD on both. Nine of the twelve false alerts come from the event rule meeting hands and zone edges, three from detection in the dark. Full tables, records and audits: the repository's README and results/.

Recording

OAK-D Pro (RGB, NV12 over USB 2), 1280×720 at 15 fps, focus fixed on the zone, exposure metered on the zone. The eval boxes (a white box and a robot-arm kit box) never appear in the training recordings, whose boxes (brown, black, gray) never appear here. The home is the author's; the person in e01–e06 is the author, and the people walking through in u01 are the author's household. About 10 GB in all (150k frames).

License and use

CC BY-NC 4.0. Please do not attempt to identify the person or the home. Cite as "EdgeGate eval sessions, chans, 2026".