schema_version string | annotator_id string | created_at string | source_document dict | annotations list |
|---|---|---|---|---|
1.0.0 | martin-lellep | 2026-04-29T22:24:40.348Z | {
"filename": "2026-04-30_hello_world.xoj",
"sha256": "34069ce04d0d053b250f6c8d2ab01d0cfa8a05b625b9b665369c0d1c6b092716"
} | [
{
"class": "word",
"page_index": 0,
"layer_index": 0,
"stroke_indices": [
0,
1,
2,
3,
4
],
"text": "Hello"
},
{
"class": "word",
"page_index": 0,
"layer_index": 0,
"stroke_indices": [
5,
6,
7,
8
],
"text": "Wor... |
Xournal++ HTR Benchmark
A benchmark dataset for evaluating handwritten text recognition (HTR) algorithms on Xournal++ documents (.xopp/.xoj files).
Dataset structure
Each benchmark sample consists of two files with a shared stem:
<stem>.xoppor<stem>.xoj— the source Xournal++ or Xournal document<stem>.gt.json— stroke-level ground truth annotations, following the schema defined in ADR 004
Ground truth is annotated at word level. Annotations reference strokes by index within the source document rather than pixel coordinates, making them usable for both image-based and future stroke-based HTR models.
Usage
This dataset is used by Xournal++ HTR to benchmark prediction algorithms.
- Downloads last month
- 35