How a Tiny Head Detector Powers Intelligent Video Conferencing
summary
A tiny YOLO-style head detector built at Cisco Norway runs <2M parameters on-device and replaces a flaky third-party person detector, enabling real-time camera framing, presence lights, and 3D meeting zones by mapping head bounding boxes to distances. The model predicts continuous 3D head orientation using a rotation matrix and geodesic loss, improving pose accuracy “for free” while slightly boosting detection; it runs with low latency on constrained hardware and unlocks reliable audio–video speaker targeting and multi‑camera cross‑view framing.
AI-written summary · about
related talks
51mHuddly - The thinking camera and applied machine learning
35mLarge-Scale Video Intelligence
49mReal Time Face Detection through a Webcam Using AI
17mGemino: Practical and Robust Neural Compression for Video Conferencing
15mC-Face: Continuously reconstructing facial expressions by deep learning contours of the face ...
21mLeveraging Python for Real-Time Image processing
16mRVEdge-Vision: A Fully Open,Ultra-Efficient On-Device AI Platform for Smart Eyewear
15mMeCap: Whole-Body Digitization for Low-Cost VR/AR Headsets