In the service of the history. AI in archivistics.
summary
A pipeline labels and recognizes faces on archival photos and ships as working prototypes, not slides. Trained on a long‑tailed set with 250k published and 37 million pending images, multi‑label models move from baseline accuracy 0.12 to about 0.58 and improve another 11% with augmentation plus a custom loss, while face recognition yields 90% precision, 65% recall, F1 ≈76%. Inference runs ~0.5s for tagging and ~1s for faces on a Titan RTX, Coral USB accelerator tests show similar speeds, and full reprocessing drops from centuries to about two years.
AI-written summary · about
chapters (20)
- 0:00AI in archivistics
- 1:35Archiving billions of photos
- 3:01Manual archiving process
- 5:19AI for automatic tagging
- 6:04Multi-label image tagging
- 8:18Baseline model results
- 10:03State-of-the-art models
- 11:34Focal loss for accuracy
- 12:41Custom loss function
- 13:45Results and comparisons
- 15:14Tagging 100 labels
- 16:20Tagging errors examples
- 18:08Face recognition in photos
- 21:05Reference photo database
- 21:49Face recognition metrics
- 22:57Face recognition surprises
- 24:26AI archiving future
- 25:11Faster processing with Coral
- 26:37Automating archival tasks
- 27:20Q&A
related talks
40m"Introduction to Face Processing with Computer Vision|"
01h 12mBuilding a live face recognition system
52mFace Recognition: From Scratch To Hatch
27mAI for social good
39mBuilding Highly Scalable Facial Recognition Pipelines
01h 00mOpening Keynote
01h 04mBuilding and Training your own Custom Image Recognition AI
36mAI-Powered Real-Time Quality Control