Unveiling the secrets of ancient coins
summary
University of Oxford’s 21 million‑object backlog becomes a quantifiable ML prototype: a heads vs tails classifier trained via Amazon SageMaker (JumpStart EfficientNet fine‑tuning), a segmentation pipeline using Open Images masks and U2‑Net / Mask R‑CNN for robust background removal and upscaling, and a feature‑vector visual search indexed in Elasticsearch. The proof‑of‑concept deploys as endpoints, returns about eight candidate matches to a query, trains in minutes on p3.2xlarge (example run ≈22 minutes), and reduces per‑coin processing from 10+ minutes or hours to a few minutes, saving an estimated three years for 300,000 coins.
AI-written summary · about
chapters (13)
- 0:00Ancient coins and AI project
- 1:18Cataloging museum collections
- 3:26Heads versus tails prediction
- 8:20Coin image processing challenges
- 10:05Image quality improvement
- 12:10Live proof of concept demo
- 15:51AWS solutions benefits
- 16:22Background removal techniques
- 18:07Building ML solution
- 22:43Segmentation model training
- 30:09Visual search engine with SageMaker
- 36:17Deploying visual search app
- 37:01Closing remarks
related talks
28mBackground removal without background knowledge
29mGet started with ML in minutes with Amazon SageMaker JumpStart
01h 29mHuman-like Visual Search Application with Small Data
27mDetecting fraud faster with Amazon SageMaker, featuring Coinbase
01h 00mOpening Keynote
55mCore ML: Vision
40mMachines are Learning
46mML Use Cases at Think Big Analytics