Segmentation evaluation and ground truth¶

Contexts select a document type from htr's profile registry: generic, printed, newspaper, manuscript, letter, ledger, or form. This selects the requested layout features and the segmentation measurement policy. The weights are uncalibrated until there are enough checked pages to compare them with transcription accuracy.

An item's Document type control overrides its context for future processing. Empty means inherit. Queued jobs retain the resolved type in their immutable context snapshot; changing an item does not change an admitted job.

Mark a checked revision¶

Save all edits, then select Ground truth… in the editor. Confirm Mark only after checking both text and layout. The action records that exact canonical revision, its original model segmentation, transcript, profile, image URL, actor, and timestamp. A stale revision fails explicitly. The same control unmarks the current revision.

Editing a marked page creates a new, unmarked revision. The previous checked snapshot stays immutable; explicitly mark the new revision when it is ready. Snapshots count toward storage quotas and are removed with the item or image.

Scribe stores only the current editable AnnotationPage, so immutable marked revisions live in annotation_ground_truth, rather than adding another editable annotation store or pretending that historical revisions already exist. ocr_runs.original_page retains the original htr geometry where available.

Export and score¶

Select Export ground truth on an item or context. The ZIP contains pages.csv, canonical IIIF sources, dimensioned htr ground-truth pages, the original segmentation, transcripts, and image URLs. No remote images are downloaded during export. An export is limited to 100 revisions and 32 MiB of source payload; select a smaller item or context when it exceeds that limit.

Extract the ZIP and run:

htr segeval-external --csv pages.csv

The CSV points at dimensioned htr JSON so scoring needs no image download and does not mistake the furthest annotation edge for the full image dimensions. The unchanged canonical IIIF page is included alongside it.

Correction telemetry and geometry¶

Saving a page computes segmentation agreement with its original OCR baseline alongside character correction distance. The component report records the profile, geometry level, calibration status, and any fallback. Handwriting profiles use baseline scoring only when real baselines exist; box-only pages report the PRImA fallback. Export-time word inference never supplies scoring ground truth.

Interactive scoring is limited to 256 elements per side, reduced to 32 for line-level pages with baselines (2048 input and 4096 sampled baseline points per side), and two concurrent evaluations per process. Larger pages remain editable and report that offline scoring is required. The complete bundle remains available for CLI evaluation.

Canonical annotations retain polygons and baselines in an SVG selector beside the authoritative xywh selector. Moving or resizing a box transforms existing points. Fresh lines have no guessed baseline. Splitting cuts the baseline at the text split and maps it into each derived box. Joining concatenates real baselines only when every input has one; the merged line uses its union box instead of inventing a polygon. Region classifications appear in the editor's Region types controls and are saved with the normal draft.