Wild-OMR: A Public Phone-Camera Benchmark for Optical Mark Recognition

Wild-OMR is a public benchmark for optical mark recognition (OMR) from phonecameras under uncontrolled capture conditions. The archive contains 769 photographs of 30 pencil-filled answer sheets, capturedwith five phones spanning seven model years under five conditions (flat, dim,glare, 30-degree tilt, hand-held/curved), plus a 300 dpi flatbed reference pass.Ground truth is exact at the level of individual bubbles: volunteers copiedpre-assigned synthetic answer keys onto anonymous, ID-numbered sheets, so theprinted key is the ground truth and no personal data was collected. Also included: the parametric sheet generator, the synthetic wild-capturegenerator with matched renderings of the same sheets, the capture protocol, fourclassical reader configurations, three learned baselines with training scripts,and the evaluation harness that produced every number in the accompanying paper. Contents- images/ 769 captures, canonical filenames- metadata.csv per-image sheet, device, condition, id provenance, SHA-256- protocol/ answer keys (ground truth), sheet geometry, capture guides- forms/ printable blank sheets and vo- code/ generators, readers, learned baselines, evaluation harness- results/ per-question predictions for per License: images and annotations CC BY 4.0; c