Twinsee at C10 Labs
Infant monitoring for Twinsee at C10 Labs — semi-supervised YOLO on pseudo-labelled video, infant posture classification, and a phone capture prototype that turns one recording into metric 3D, depth and camera trajectory
ML/CV Graduate Co-op at C10 Labs, Boston, on Twinsee, a camera-based infant monitor (Sep 2026 – present).
- Semi-supervised detection — YOLO trained on video frames pseudo-labelled by Grounding DINO, YOLO and ByteTrack: precision 38% → 93%.
- Infant posture classification — from DINO features, infant pose, and the gravity direction estimated by GeoCalib.
Phone capture prototype
How much can one phone recording give, before any extra hardware? An Android recorder on a Samsung Galaxy S25 captures wide + ultrawide video (1080p30 each, stabilization and distortion correction off, exposure locked) with every motion sensor on the same clock, plus the factory calibration and per-frame camera metadata. Everything below is computed offline from one 16-second orbit of a glass cup on a desk — a deliberately hard scene: transparent glass, a glossy black desk and monitor, motion blur.
The full demo report is below: what the phone records (sensor and camera dashboard), then gravity, north and height; the metric SfM cloud; planes and object measurement; stereo + Depth Pro depth fusion and its accuracy; a Gaussian splat; a MapAnything dense cloud; and the MegaSaM trajectory fixed with phone scale. The 3D viewers are live: drag to orbit, scroll to zoom.