From 87d2d145b992ab52670288a20da15f2d5a016f94 Mon Sep 17 00:00:00 2001 From: Robert Sachunsky Date: Tue, 6 Jun 2023 12:21:23 +0200 Subject: [PATCH] ocrd_eval: add alternative order metrics --- ocrd_eval.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/ocrd_eval.md b/ocrd_eval.md index 7d82d68..162049a 100644 --- a/ocrd_eval.md +++ b/ocrd_eval.md @@ -374,6 +374,13 @@ The success measure is then given by $s = \frac{1}{e * (1/e_{50}) + 1}$ +Furthermore, in the special case where… +- the segmentation on both sides is identical (only order differs → no matching needed) +- order is always strictly monotonic (no unordered subgroups involved → no misclassification other than successor vs predecessor) + +…[Quirós and Vidal 2022](https://ieeexplore.ieee.org/document/9413256) evaluate reading order simply as rank distance, +using the metrics _Normalized Spearman's footrule distance_ and _Kendall's Tau rank distance_. + ##### mAP (mean Average Precision) This score was originally devised for object detection in photo scenery (where overlaps are allowed and cannot conflict with text flow). @@ -542,6 +549,7 @@ See [OCR-D workflow guide](https://ocr-d.de/en/workflows#evaluation). * * Reading Order Evaluation: * + * * More background on evaluation of OCR *