Core ML: Vision
summary
A Core ML Vision demo builds an object detector for six-sided dice using 45 images, 6 classes and transfer learning (Apple suggests ~5,000 iterations; the demo uses ~1,000 and runs overnight). Annotation is automated with an OpenCV annotation script and supplemented by Roboflow, producing JSON boxes; evaluation shows confidences like 100%, 92%, 88% depending on angle and lighting. Model runs as on-device inference with tunable confidence thresholds and needs ~50+ varied examples for stability.
AI-written summary · about
chapters (28)
- 0:00Intro
- 2:59Computer Vision
- 3:49Image Identification
- 4:23Controlling Processes
- 5:11Detecting Events
- 5:49Interaction
- 6:22Navigation
- 7:27Medicine
- 9:00Machine Vision
- 10:31Autonomous Vehicles
- 11:56Engaging ordinary people
- 14:36Vision Framework
- 15:02Vision Framework Demo
- 16:51Creating a Project
- 17:31Training Data
- 19:19Image Detection
- 20:40Number of iterations
- 21:46Grid size
- 23:51Training
- 25:57Annotation
- 27:01JSON
- 28:40Automate
- 30:02Roboflow
- 32:48Scaling Factor
- 33:49Python Script
- 35:41Annotations
- 40:00Snapshot
- 42:02Model
related talks
32mReal-Time Object Detection with CoreML
40mNavigating the Machine Learning Landscape for Apple Devices
30mLearning CoreML with fun
40mDeep Learning on iOS
34mDemystifying ML in iOS
33mComputer Vision, Not Just for Breakfast Anymore
49mKeynote: From Zero to ML on Google Cloud Platform
45mMobiconf 2018: Meghan Kane, Novoda