411k cut-outs from Britannica: 29M param model segments historical illustrations

vanstriendaniel · x · 2026-08-27

A dataset of 411,385 instance masks for illustrations from Encyclopaedia Britannica (1768–1929) was released. The project used a Roboflow RF-DETR segmentation model (29M params), fine-tuned by a coding agent on Hugging Face with labels bootstrapped from Falcon-Perception. Running at 1s/page on a laptop CPU with a total cost of $2, it demonstrates an efficient workflow for large-scale historical document processing using small models.

Original post →

More from coding & agent

coding & agent channel →