Skip to content

Test data notes

Jump to bottom
Dan Vanderkam edited this page Jun 10, 2026 · 8 revisions

  • New Orleans 1951 Vol 5

    • p434: OCR misses "S. FRONT" and misreads a scribble as "MIRO" which produces a 1-gcp fit miles away.
    • p454: Napoleon Ave is a split highway, which produces a rotated fit in this case.
    • p436: 1-gcp along the water, this isn't so bad.
    • p437: this looks like a decent fit, maybe ever so slightly rotated
    • p433: 1-gcp fit, slightly rotated from where the roads used to go
    • p438: not too bad, angle is just slightly off

  • Detroit 1929 Vol 11

    • p8: 1-gcp based on a misidentifcation of the “M P” in “COMPANY” as “DEAN”. The renamed CONNORS/CONNER prevents additional intersections from saving us.
    • p50: a correct 1-gcp that’s oddly rotated, maybe because E Jefferson turns at this intersection.
    • p93: another correct 1-gcp that's oddly rotated; Kelson Drive and Avondale Street both turn a bunch in OSM but are straight in the Sanborn map.
    • p28: the streets have changed here in a way that rotates the image. I’m surprised any of these streets are inliers. There is a “Lane Street” in Detroit that creates problems for all other Lanes.
    • p5: This is because of St. Jean vs. Old St. Jean Ave.
    • p44: Lakewood Street jogs across E. Jefferson Ave but a poor fit isn't penalized, probably because of the 500ft extrapolation.
    • p71: Also an E. Jefferson issue.
    • p43: Newport has a slight jog across Kercheval that shifts some streets.
    • p30: 1-gcp on E. Jefferson, but no “jog” here. The issue is that it’s “Connors Av” on the Sanborn map but “Conner Street” in OSM. Unclear if this was a name change or a mistake in Sanborn, but it throws this off a bit.
    • p77: Good 1-gcp fit, just slightly disagree w/ OIM on scale.

  • Chicago 1950 Vol 1

    • p105W: there’s a bunch of labels on the streets (Pennsylvania, Western) that get detected instead of the street name (TAYLOR). There’s also “BRANCH of CHICAGO RIVER” that are all misdetected. And ELLSWORTH is detected, but it’s written 90° off. So we get a bad 1-gcp that’s 3 miles off.
    • p29N: comes out rotated, penalizing rotations (rather than a strict inlier/outlier fraction) might help here. This is pathologically O(N^2).
    • p106W: Misreads "TRUSS" as "CANAL" and uses this to come up with a 1-gcp fit that is, surprisingly, in roughly the right spot and is correctly-rotated. This should be dropped.
    • p89W
    • p92W
    • p42W
    • Misses:
      • p1N: possible; misidentifies "W CHICAGO AV" as "S CHICAGO AV" due to a number. Gets N LARRABEE STREET with very low confidence (0.002) due to trailing "(ROBERTS)". Recognizes ERIE but with low confidence (0.061).
      • p41N: thrown out as a scale mismatch; I think the location is correct, the scale isn't that bad; the Sanborn map is more schematic than the OSM streets.
      • p50N: thrown out as a scale mismatch; our fit is rotated. E ILLINOIS x ST CLAIR cross but we miss the intersection, I think because it's "East Lower Illinois Street" in OSM, and the "Lower" throws things off.
      • p51N: I don't think SENECA street still exists here.
      • p53N: Lots of bad detections (MICHIGAN CANAL) but misses E NORTH WATER. Happily thrown out as a scale mismatch.
      • p54N: There's only one street here (E NORTH WATER), so this would be hard.
      • p55W: Thrown out as a scale mismatch. It's not so bad. W CHICAGO x N. HALSTED would be a better choice of GCP, why doesn't it pick that?
      • p61W: I think it makes different choices about E vs. W streets for the two intersections, which results in a scale mismatch. It could consider this in making its choice about which street is which. (There are also lots of misdetected streets.)
      • p98W
      • p99W

  • Champaign, Ill. 1915

    • p13__1: tiny map. mapsnap's GCPs are good, but it's unclear to what extent the streets are drawn to scale here. OIM's fit is based on parcels.
    • Misses:
      • p2__2: very small, would have to be 1-gcp. NEIL is detected, but BEARDSLEY AVE is only detected with 0.1 confidence.
      • p3__3: very small, there's only one street (N MARKET) which is detected.
      • p4__2: large, but only one street (VICTOR) which is detected.
      • p4__3: small, would have to be 1-gcp. HARRIS is detected, but BEARDSLEY is 0.07 confidence and rotated 90° from the expected orientation.
      • p4__4: small, could be 1-gcp with N ELM and EUREKA. Both are detected, but N ELM is rotated 90° from the expected orientation.

  • New Orleans 1896 Vol 2

    • p169: this is a false positive, Mapsnap's fit is better than OIM's.
    • p157: 1-gcp that doesn't quite get the angle right. The streets have changed a lot here due to highways.
    • p149: this is a false positive, Mapsnap's fit is better than OIM's.
    • p177: mapsnap makes a poor choice of intersections because South Roman Street "jogs" now in a way that it didn't in 1896.
    • p171: this is a false positive, Mapsnap's fit is better than OIM's.
    • p148: this fit looks pretty good, maybe a tiny bit rotated

  • Brooklyn, NY 1939 Vol. 2

    • p13: Misses the only two vertical streets, PIERREPONT and CLARK, due to CRAFT picking up unrelated ink next to them. Since all of the surrounding pages are referenced successfully, this might be a good test case for #10.

Clone this wiki locally