RaphaelBfr/morphology4metrology

Leveraging Morphology for Historical Script Metrological Analysis

Pretrained checkpoint for the architecture presented in Leveraging Morphology for Historical Script Metrological Analysis (ICDAR 2026).

Authors: Malamatenia Vlachou Efstathiou*, Raphaël Baena*, Dominique Stutzmann, Mathieu Aubry

Links

Paper (arXiv)	arxiv.org/abs/2606.09446
Project page	malamatenia.github.io/morphology4metrology-analysis
Code	github.com/malamatenia/learnable-DTLR
Metrological analysis	github.com/malamatenia/morphology4metrology-analysis
Paper dataset	Zenodo — BnF fr. 2813 (DOI: 10.5281/zenodo.18745702)

Description

This checkpoint is the pretrained detector used as starting weights for step 0 of the training pipeline described in the paper.

The architecture learns interpretable, learnable character prototypes along with precise, instance-level bounding boxes using only line-level transcription supervision.

Note: This Hugging Face repository hosts the pretrained weights only. Installation, training, and inference are handled in learnable-DTLR. To perform the downstream metrological and paleographical analysis (and reproduce the visualizations reported in the paper), export the outputs from this pipeline and use morphology4metrology-analysis.

Files in this repository

File	Description
checkpoint.pth	Pretrained detector — use as --model_checkpoint_path in step 0 (with --init)
config/Latin_accent.py	Model config — use as --model_config_path

Download

pip install huggingface_hub
hf download RaphaelBfr/morphology4metrology checkpoint.pth --local-dir ./weights
hf download RaphaelBfr/morphology4metrology config/Latin_accent.py --local-dir ./weights/config

Or in Python:

from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download("RaphaelBfr/morphology4metrology", "checkpoint.pth")
config = hf_hub_download("RaphaelBfr/morphology4metrology", "config/Latin_accent.py")

Quick start

1. Clone and install learnable-DTLR.
2. Download the weights above into ./weights/.
3. Prepare a line dataset under datasets_path (see Input data format).
4. Run the 3-step training pipeline starting from this checkpoint.

You do not need to run synthetic pretraining — download this checkpoint and go directly to step 0.

Training pipeline

Step	Goal	Input	Output
0	Frozen boxes — learn prototypes + classifier	checkpoint.pth + --init	model.pth, reconstructor.pth
1	Full training — boxes + prototypes jointly	step 0 outputs	model.pth, reconstructor_unfrozen.pth
2	Per-document or per-script finetuning	step 1 outputs	one checkpoint folder per subset

Ready-made scripts for the paper dataset: scripts/btv1b84472995/.

Step 0 — frozen boxes, learn prototypes + classifier

python reconstruction.py \
  --dataset_file dataset \
  --data_folder btv1b84472995 \
  --space_index 0 \
  --model_config_path config/Latin_accent.py \
  --max_e 20 \
  --num_fine_classes 2 \
  --step 0 \
  --batch_size 16 \
  --sprite_size 32 \
  --line_resize_h_ref 90 \
  --line_resize_max_width 1400 \
  --init \
  --tag btv1b84472995_step_0 \
  --loss L1 \
  --model_checkpoint_path ./weights/checkpoint.pth

Argument	Role
--step 0	Freeze detector boxes; train reconstructor + classification head only
--init	Rebuild classifier for the dataset charset from the pretrained checkpoint
--model_checkpoint_path	This checkpoint (checkpoint.pth)
--tag	Run name; outputs go to logs_reconstruction/<tag>/

Outputs (logs_reconstruction/btv1b84472995_step_0/): model.pth, reconstructor.pth

Step 1 — full model

python reconstruction.py \
  --dataset_file dataset \
  --data_folder btv1b84472995 \
  --space_index 0 \
  --model_config_path config/Latin_accent.py \
  --max_e 100 \
  --num_fine_classes 2 \
  --step 1 \
  --batch_size 8 \
  --sprite_size 32 \
  --line_resize_h_ref 90 \
  --line_resize_max_width 1400 \
  --learning_rate 1e-4 \
  --weight_loss_reconstruction 3 \
  --tag btv1b84472995_step_1 \
  --loss L1 \
  --model_checkpoint_path logs_reconstruction/btv1b84472995_step_0/model.pth \
  --reconstructor_path logs_reconstruction/btv1b84472995_step_0/reconstructor.pth

Outputs (logs_reconstruction/btv1b84472995_step_1/): model.pth, reconstructor_unfrozen.pth, sprites/, transcribe.json

Step 2 — per-document finetuning

python reconstruction.py \
  --dataset_file dataset \
  --data_folder btv1b84472995 \
  --documents \
  --space_index 0 \
  --max_e 140 \
  --num_fine_classes 2 \
  --step 2 \
  --sprite_size 32 \
  --line_resize_h_ref 90 \
  --line_resize_max_width 1400 \
  --mask_sprite \
  --learning_rate 1e-2 \
  --learning_rate_background 1e-5 \
  --batch_size 8 \
  --tag btv1b84472995_finetune \
  --output_dir logs_reconstruction/ \
  --prototypes_only_path logs_reconstruction/btv1b84472995_step_1/reconstructor_unfrozen.pth \
  --model_checkpoint_path logs_reconstruction/btv1b84472995_step_1/model.pth \
  --annotation_file /path/to/datasets/btv1b84472995/annotation.json

Step 2 also supports --script <name> to finetune per hand/style label instead of --documents.

Input data format

<datasets_path>/<data_folder>/
├── annotation.json
├── images/
│   ├── <document_prefix>/
│   │   ├── <line_id>.png
│   │   └── ...
│   └── ...

{
  "IB15304245v_eSc_line_65f6c260.png": {
    "label": "sy enlas armas por bueno q̃ sea",
    "page": "IB15304245v.jpg",
    "split": "train",
    "script": "Southern_Textualis"
  }
}

Set the datasets root in datasets/config.json:

{
  "datasets_path": "/path/to/your/datasets"
}

Export paleography_input/

After step 1 and finetuning, export sprites and bbox JSONs for downstream analysis:

python export_paleography_input.py \
  --step1_dir logs_reconstruction/btv1b84472995_step_1 \
  --finetune_dir logs_reconstruction/btv1b84472995_finetune \
  --dataset_file dataset \
  --data_folder btv1b84472995 \
  --model_config_path config/Latin_accent.py \
  --line_resize_h_ref 90 \
  --line_resize_max_width 1400

Then use morphology4metrology-analysis on the exported paleography_input/ folder.

Paper dataset

Experiments use the Grandes Chroniques de France line dataset (Paris, BnF, fr. 2813), published on Zenodo. Download dataset.zip, extract under your datasets_path, and use --data_folder btv1b84472995.

Citation

@inproceedings{vlachou2026metrology,
  title     = {Leveraging Morphology for Historical Script Metrological Analysis},
  author    = {Vlachou-Efstathiou, Malamatenia and Baena, Raphael and
               Stutzmann, Dominique and Aubry, Mathieu},
  booktitle = {Document Analysis and Recognition -- ICDAR 2026},
  publisher = {Springer},
  year      = {2026}
}

Preprint: arXiv:2606.09446