Knitted Disruption
summary
Crowdsourced hundreds of hand‑knit photos, labeled with SageMaker Ground Truth and trained with SageMaker image classifiers, then deployed behind a React front end; scope is reduced to stockinette or not rather than stitch‑by‑stitch reverse engineering. After image preprocessing (grayscale, rotations, added non‑knit clutter) and ~15 training runs, the best model reaches nearly 95% accuracy; total cloud training spend is about $25 training cost. Data labeling involved 14 volunteers over ~15 hours.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:06I'm a knitter
- 2:58Embed data
- 7:27Knitting machines
- 15:45Benefits of a DSL
- 21:07Could a computer READ knitting?
- 21:57Reverse engineering
- 28:42Get training data
- 33:15Label training data
- 36:14Train ML model ml.p3.2xlarge
- 37:08Deploy ML model
- 37:18Build web frontend for inference
related talks
40mThe Effortless Development of Custom Computer Vision Models
46mImage Recognition with Deep Learning From prototype to production
22mCreating high-quality training sets for machine learning models
26mKnitting for Javascripters
28mBackground removal without background knowledge
01h 00mOpening Keynote
58mImproving disaster response with machine learning
29mHow the NFL builds computer vision training datasets at scale