The Effortless Development of Custom Computer Vision Models
summary
A sign-language classifier is built from 764 images (12 labels) by stitching nine frames into a 3x3 grid and training with ResNet-18 transfer learning on SageMaker (148 training jobs, 49 GPU hours, ~83% accuracy). The same dataset is retrained in about an hour on Amazon Rekognition Custom Labels (45 minutes training) with slightly better macro F1/precision/recall. Rekognition auto-generates manifest files, serves via managed endpoints with adjustable assumed thresholds per label, and recommends starting endpoints at minimum TPS = 1.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 2:11Simplifying the problem
- 3:55Building the model in SageMaker
- 9:13Introducing Amazon Rekognition Custom Labels
- 11:59Comparing SageMaker and Rekognition
- 21:45Understanding evaluation metrics
- 34:21Deploying and using the model
- 37:31Choosing between SageMaker and Rekognition
- 39:56Closing and resources
related talks
29mRekognition Live Demo
01h 00mOpening Keynote
35mHow deep learning can help interpret American Sign Language
01h 04mBuilding and Training your own Custom Image Recognition AI
30mBring AutoML to all users across your business
24mImage Recognition for safety on the factory floor
40mMachines are Learning
01h 01mGetting Started with Computer Vision