One real phone photo of a Japanese math worksheet. Every result scored by the same
metric, with the diff showing exactly what the metric counts.
Print recall (safety, want 1.00) = fraction of printed strokes still present.
Handwriting erase (utility) = fraction of handwriting strokes removed.
Robust by design: a pixel is "a mark" if it's
darker than its local paper
(median-blur background) — not absolute black/white — so warm/shadowed paper doesn't break it
(the bug that scored the ideal clean 0.42). Reduction-normalized → exact baselines
(
input → 1.00 / 0.00,
ideal → 1.00 / 1.00). Result is ORB+homography-registered to
the input first, so a few-px shift isn't mistaken for erasure.
handwriting removed (good)
handwriting left behind
PRINT erased (catastrophic)