Orli

Orli (ordered regression of lines) is a layout-analysis model that detects text-line baselines and emits them directly in reading order. It is designed for historical-document OCR pipelines that need explicit line geometry without a separate reading-order heuristic.

The method is described in the article End-to-End Text Line Detection and Ordering.

Installation

pip install .

Orli integrates with kraken 7 through its model plugin system.

Model

The release base model is trained on 200000 pages spanning ten writing systems. It is published available through HTRMoPo with DOI 10.5281/zenodo.20558179.

Download it with kraken:

kraken get 10.5281/zenodo.20558179

The command prints the model directory and the downloaded model file, necessary for fine-tuning and programmatic inference.

Run baseline segmentation with kraken. This example writes PAGE XML:

kraken --precision bf16-mixed -i input.jpg output.xml -x segment -bl --model orli_base.safetensors

The base model only works in bfloat16 precision! Other precisions are likely to cause runaway generation.

Programmatic inference uses the complete model path printed after download:

from PIL import Image
from orli.pred import segment

im = Image.open("input.jpg")
segmentation = segment(im, "/path/to/kraken/download/orli_base.safetensors")

Scores

Line metrics are computed using the cBAD evaluation score implemented in orli test. Footrule is normalized Spearman footrule, where lower is better.

Test Set

Model	Precision	Recall	F1	Cov.	Footrule	Kendall tau
base	0.9554	0.9564	0.9559	0.9667	0.0304	0.9649

cBAD 2019

Model	Precision	Recall	F1	Cov.	Footrule	Kendall tau
base	0.9378	0.9302	0.9340	0.9406	0.0768	0.9113
fine-tuned	0.9395	0.9306	0.9351	0.9421	0.0720	0.9165

Reading-Order Benchmarks

Dataset	Model	Precision	Recall	F1	Cov.	Footrule	Kendall tau
OHG	base	0.9940	0.9937	0.9938	0.9993	0.0033	0.9967
FCR	base	0.9894	0.9874	0.9884	0.9905	0.0028	0.9971
ABP	base	0.8505	0.7919	0.8201	0.8071	0.5372	0.2878
ABP	fine-tuned	0.8498	0.7806	0.8137	0.7931	0.0898	0.8972

Dataset Preparation

Orli trains on Arrow datasets compiled from PageXML or ALTO files. The compiler stores each line as a normalized baseline polyline in source-file order. The arrow files are NOT compatible with kraken's compiled datasets.

orli compile -o dataset.arrow --allow-textless *.xml

For large images, pre-resizing during compilation reduces training I/O. The base model uses a high-resolution input size of 1920x1440:

orli compile -o dataset.arrow --allow-textless -r 1920 1440 *.xml

Compilation uses the implicit reading order, i.e. the sequence of line elements in the source file. Other reading-order annotations are ignored.

Training and Fine-Tuning

Training and fine-tuning are configured either through command-line options or a YAML file. For fine-tuning the released base model, keep the high-resolution input size and load the downloaded orli_base.safetensors file:

precision: bf16-mixed
device: auto
num_workers: 12
num_threads: 1
train:
  training_data:
    - train.arrow
  evaluation_data:
    - val.arrow
  checkpoint_path: experiments/orli_finetuned
  image_size: [1920, 1440]
  optimizer: AdamW
  epochs: 8
  lrate: 5e-5
  weight_decay: 1e-4
  schedule: cosine
  cos_t_max: 8
  cos_min_lr: 1e-5
  warmup: 1000
  augment: true
  batch_size: 8
  val_batch_size: 8
  accumulate_grad_batches: 8
  baseline_num_points: 16

orli --config finetune.yaml train --load "$MODEL"
orli --config finetune.yaml train --resume /path/to/checkpoint.ckpt

The training command writes the best checkpoint and converts it to safetensors automatically. The resulting best_*.safetensors file can be used with kraken segment in the same way as the base model.

Evaluation

Evaluate a model on an Arrow dataset with baseline detection metrics and reading-order metrics:

orli test --load model.safetensors test.arrow

Citation

@misc{kiessling2026orli,
  title = {End-to-End Text Line Detection and Ordering},
  author = {Benjamin Kiessling},
  year = {2026},
  eprint = {2606.04166},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2606.04166}
}

License

Orli is released under the Apache License 2.0.