evaluation_kit/REPORT_JSON_METRICS.md

Report JSON Metrics

This document explains every field emitted in an overall_accuracy_report.json produced by this kit (via evaluate_from_hf.py or evaluation/measure_overall_accuracy.py).

The report is generated by, e.g.:

python3 evaluate_from_hf.py \
  --dataset RLALT/ACoPPer \
  --predictions-dir path/to/your/evaluation_csvs \
  --output-dir results/

See README.md for the full pipeline explanation. The metrics below are computed primarily in evaluation/measure_accuracy.py and aggregated in evaluation/measure_overall_accuracy.py.

Conventions

• Row: one predicted OCR row (group_row) composed of multiple OCR word boxes.
• Word box: one OCR detection box from the CSV (one row in the CSV file).
• Annotation box / GT box: one labeled text box from the JSON annotations.
• Region: a connected component of transcribed GT boxes according to which boxes are jointly touched by predicted rows. Isolated GT boxes also become one-box regions, even when no predicted row is assigned.
• CER: character error rate, computed as char_edit_distance / gt_char_count with safe_error_rate behavior (see below).

Rounding / safe helpers

• safe_rate(count, total): round(count / total, 6) (or 0.0 when total == 0).
• safe_mean(values): round(sum(values) / len(values), 6) (or 0.0 when empty).
• safe_error_rate(edit_distance, gt_len):
  • if gt_len == 0: returns 0.0 when edit_distance == 0, else 1.0
  • else: round(edit_distance / gt_len, 6)

Top-Level JSON Shape

The file is a JSON object with these keys:

• summary: aggregate metrics across all matched page pairs.
• pages: list of per-page reports.
• split_line_groups: all split-line groups across pages (already enriched with page metadata).
• ignored_rows: all ignored watermark rows across pages (already enriched with page metadata).
• excluded_labels: report-level label metadata used by filtered metrics. Written once at the end of the JSON object.

summary (Aggregate Metrics)

All fields below exist under summary at the top level. The same set (minus pair_count) exists under each pages[i].summary.

• pair_count: number of matched (csv,json) page pairs processed.
• total_rows: number of predicted rows evaluated (excluding ignored watermark rows).
• ignored_watermark_rows: number of predicted rows ignored because they were classified as watermark.
• exactly_one_box: count of predicted rows whose final status is exactly_one_box. Split-line postprocessing runs before this count, so rows reclassified as split_line are not included.
• exactly_one_box_rate: exactly_one_box / total_rows.
• multiple_boxes: count of predicted rows whose final status is multiple_boxes.
• multiple_boxes_rate: multiple_boxes / total_rows.
• no_box: count of predicted rows whose final status is no_box.
• no_box_rate: no_box / total_rows.
• split_line: number of split-line groups detected. Each group corresponds to one GT box whose text was split across multiple predicted row fragments.
• split_line_rate: split_line / total_rows. This is group count over row count, so compare it carefully with split_line_rows_rate.
• split_line_rows: number of predicted rows whose final status is split_line.
• split_line_rows_rate: split_line_rows / total_rows.
• mean_best_coverage: mean of dominant_coverage across evaluated rows (after watermark removal). Higher is better.

OCR region metrics (ocr_region_*)

These are computed over ocr_regions (regions are connected components of transcribed GT boxes).

• ocr_region_count: number of OCR regions.
• multibox_region_count: number of OCR regions containing more than one GT box.
• ocr_region_mean_cer: simple mean of per-region CER (mean(cer_i)).
• ocr_region_cer: corpus CER across all regions, computed as (sum char_edit_distance) / (sum gt_char_count).
• ocr_region_cer_buckets: compact CER bucket summary. Each bucket has count and rate, where rate = count / ocr_region_count.

CER buckets:

• lt_0_1: cer < 0.1
• 0_1_to_0_3: 0.1 <= cer < 0.3
• 0_3_to_0_6: 0.3 <= cer < 0.6
• 0_6_to_1: 0.6 <= cer <= 1.0
• gt_1: cer > 1.0 (can happen when the predicted string has enough insertions that character edit distance exceeds GT character count).

Normal single-box region metrics

normal_single_box_region is computed over the subset of ocr_regions where the region is considered a "clean" box match:

• len(box_ids) == 1 (region corresponds to exactly one GT box)
• the region has at least one assigned OCR row internally
• every assigned OCR row has final status == "exactly_one_box"

This filter excludes:

• merged/multi-box regions
• split-line rows
• regions built from rows that end up multiple_boxes or no_box

Fields inside normal_single_box_region:

• count: number of regions that satisfy the filter.
• mean_cer: simple mean of CER across filtered regions.
• cer: corpus CER across filtered regions.

Important: this is still a text OCR metric, not a pure layout metric. The filter says column/box separation worked cleanly; CER can still be high if recognition, line ordering, punctuation, or hyphenation failed inside the correct GT box.

Optional filtered region mode

By default, measure_accuracy.py and measure_overall_accuracy.py compute only the unfiltered ocr_region_* metrics. To evaluate filtered regions, pass a comma-separated --filter value, for example:

• --filter non-armenian
• --filter graphics
• --filter non-armenian,graphics

When filters are active, the primary ocr_region_* fields are computed from filtered regions. Excluded GT boxes are removed from region GT text, and predicted word fragments whose centers are inside excluded boxes are removed from region predicted text.

Supported filters:

• non-armenian: excludes GT text boxes where more than 90% of letters are Latin or Cyrillic. The ratio ignores digits, punctuation, and spaces.
• graphics: excludes boxes labeled Graphics.
• photo: excludes boxes labeled Photo.
• image: excludes image-related labels: Photo, Graphics, SealFigure, FrontPicture.
• header: excludes header/title-like labels: Headline, Kicker, Banner, Deck, Subhead, Nameplate, Masthead, FrontStory.
• image-header: excludes both image-related and header/title-like labels.

When filters are active, summary.filter contains:

• filters: canonical filter names that were applied.
• text_box_count: number of transcribed GT boxes.
• excluded_box_count: number of transcribed GT boxes excluded by the active filters.
• excluded_box_rate: excluded_box_count / text_box_count.
• included_box_count: transcribed GT boxes remaining after exclusion.
• threshold: present when non-armenian is active.
• labels: present when label-based filters are active.

Per-page and aggregate reports also include filtered_text_boxes, the excluded boxes with box_id, labels, normalized GT text, matched filters, and non-Armenian ratio fields.

Empty-word-in-GT-box metric (missing_text_*)

This measures a detection/recognition failure mode: the OCR CSV contained empty-text word boxes inside GT text areas.

• total_detected_word_boxes: total number of OCR word boxes in the CSV(s), including boxes whose text is empty.
• missing_text_boxes: number of OCR word boxes whose text is empty and whose center lies inside a GT box with non-empty transcription text.
• missing_text_box_rate: missing_text_boxes / total_detected_word_boxes.

pages[i] (Per-Page Report)

Each element in pages has:

• page_name: stem name used to match page_name.csv with page_name.json.
• predictions_csv: path to the CSV used for that page.
• annotations_json: path to the annotation JSON used for that page.
• summary: per-page summary metrics (same meaning as top-level summary, but page-scoped and without pair_count).
• ocr_regions: list of region records (see below). Aggregate reports keep all region records so downstream visualizations can use the report JSON as the region source.
• split_line_groups: list of split-line groups for that page (see below).
• ignored_rows: list of ignored watermark rows for that page (see below).

pages[i].ocr_regions[j] (Region Record)

Keys:

• region_id: stable identifier for the region (currently the first GT box id in the ordered region).
• box_ids: ordered GT box ids that belong to this region.
• box_types: ordered GT box types, derived from the first rectangle label for each GT box, in the same order as box_ids; null when a box has no label.
• gt_boxes: list of GT box rectangles for the region in [x_min, y_min, x_max, y_max] format (same order as box_ids).
• gt_box_details: ordered GT box metadata. Each item has box_id, box_type, labels, and box.
• predicted_box: merged bounding box of the predicted lines used for this region in [x_min, y_min, x_max, y_max] format (or null if none).
• predicted_line_count: number of reconstructed predicted lines.
• normal_single_box_region: boolean flag used by the summary normal_single_box_region metrics. It is true only for one-GT-box regions with at least one assigned row and only final exactly_one_box rows.
• text_metrics: text-comparison metrics for normalized GT text vs normalized predicted text (see below).

text_metrics (in region records)

Keys:

• gt_normalized_text: normalized GT text, with raw report text hidden.
• pr_normalized_text: normalized predicted text, with raw report text hidden.
• gt_char_count: len(gt_normalized_text).
• predicted_char_count: len(pr_normalized_text).
• char_edit_distance: Levenshtein distance over normalized strings.
• cer: safe_error_rate(char_edit_distance, gt_char_count).

split_line_groups (Per-Page and Aggregate)

A split-line group indicates that multiple predicted rows are fragments of the same GT box line and were reclassified to split_line.

Keys:

• group_id: stable group id.
• box_id: GT box id the split-line group belongs to.
• box_text: GT text for that box.
• fragment_count: number of row fragments in the group.
• row_ids: predicted row ids in reading order.
• row_texts: the predicted text per row id.
• page_name, predictions_csv, annotations_json: present in the aggregate split_line_groups list (and in per-page lists after enrichment).

ignored_rows (Per-Page and Aggregate)

Rows ignored as watermark.

Keys:

• row_id: predicted row id.
• row_text: predicted row text.
• words: list of word texts in the predicted row.
• reason: currently "watermark".
• page_name, predictions_csv, annotations_json: present in the aggregate ignored_rows list (and in per-page lists after enrichment).

Generating These Reports

Against the RLALT/ACoPPer or RLALT/ACoPDoc Hugging Face dataset (see README.md):

$ python3 evaluate_from_hf.py \
    --dataset RLALT/ACoPPer \
    --predictions-dir path/to/your/evaluation_csvs \
    --output-dir results/

This writes the standard four report files to --output-dir:

overall_accuracy_report.json
overall_accuracy_filtered_non_armenian_report.json
overall_accuracy_filtered_graphics_headers_images_photos_report.json
overall_accuracy_filtered_all_report.json

The overlay PNGs and matching JSON legends are written under region_overlays/<report-name>/ inside the selected results directory.