BIRDEYE — On-device baby bassinet sleep-state cascade

BIRDEYE is a 3-stage MobileNetV3 cascade trained for BILBO, a self-hosted baby bassinet monitor. It classifies bassinet frames into sleep state on-device, falling back to a cloud vision API only on the ~1% of frames the cascade can't decide.

Files

File	Stage	Architecture	Input	Output
presence_classifier.pt	1: presence	MobileNetV3-Small	full bassinet crop	present / not_present
face_detector.pt	2: face detect	MobileNetV3 (trainable detector)	full bassinet crop	bbox + confidence
eye_state_classifier.pt	3: eye state	MobileNetV3-Small	face crop (448x448)	eyes_open / eyes_closed
meta.json	metadata	—	—	deployment version, crop sizes

Version: v_20260430_165141 — the production version on the maintainer's BILBO deployment as of 2026-05-26.

How the cascade runs

frame -> presence_classifier
          |- not_present -> done
          +- present -> face_detector
                         |- low-conf -> cloud-API fallback
                         +- bbox    -> eye_state_classifier (face crop)
                                       +- eyes_open / eyes_closed -> state smoother

Eye-state crop is 448x448 (see meta.json). All classifiers are MobileNetV3-Small for fast CPU inference (~0.16s end-to-end on an M1 Mac mini).

Intended use

Drop-in weights for BILBO. Place under pipeline/models/v_<timestamp>/ and symlink pipeline/models/latest to that directory. BILBO's capture container hot-reloads on the next tick.

Training

Trained on labeled bassinet frames from the maintainer's own deployment via the BILBO retraining loop. Not trained on any public dataset. Tuned to a single camera angle and lighting environment — your mileage may vary if you point a different camera at a different bassinet. Re-train on your own corrections via the loop documented in BILBO's training docs.

Fallback face detector (not included)

The BILBO pipeline can fall back to OpenCV's YuNet ONNX face detector if face_detector.pt returns low confidence. YuNet is Apache-2.0 and ships separately — grab it from opencv-zoo and drop it at pipeline/models/face_detection_yunet_2023mar.onnx.

License

MIT, matching the parent BILBO project. Use freely for personal or commercial projects.

Caveats

• Single-environment training set. Biased toward the maintainer's specific bassinet, camera, and lighting.
• Personal project; no maintenance commitment. New versions uploaded ad-hoc as BILBO is retrained.
• BILBO itself defaults to GPT-4o as a cloud-fallback for low-confidence frames; on its own, this cascade should be combined with a fallback or with manual review.