Huddly - The thinking camera and applied machine learning
summary
Huddly shows a USB collaboration camera that runs convolutional neural nets on-device using a hardware-accelerated CNN (Myriad X), a high-resolution sensor and microphone array, with real prototypes built and shipped. The team labels recorded meetings, trains SSD/YOLO-style detectors in TensorFlow/Caffe (training ~20–24 hours), and tunes a cost function balancing framing accuracy versus motion, evaluated by an auto-zoom scoring server plus live user tests and CI gating. The feature ships with the product.
AI-written summary · about
chapters (8)
- 0:00Start
- 9:38Image Captioning
- 13:13Edge Computing and Why Are We Relevant
- 19:30Auto Zoom Experience
- 20:16Workflow
- 27:15Live User Testing
- 37:08Manual Controlling
- 49:51Person Detection
related talks
47mHow a Tiny Head Detector Powers Intelligent Video Conferencing
17mHow Cephable Applies On-Device AI to Accessibility
56mTensorFlow.js: Make a smart webcam in JS with a pre-trained ML model
36m"Lights, camera, AI!" - Automated sports videography
20mScaling Edge AI with Elixir and Nerves
46mML Use Cases at Think Big Analytics
34mTesting Live Video Streams At Scale
33mSee More, Be More