regnetx-8gf-polyline-hebrew-samaritan

Hebrew + Samaritan polyline baseline detector: RegNetX-8GF + D-FINE HybridEncoder + PolylineTransformer.

Hub repo: johnlockejrr/regnetx-8gf-polyline-hebrew-samaritan
Primary file: best_cbad_f1.safetensors (~208 MB)
Paper (Zenodo preprint): doi:10.5281/zenodo.22059333
Recommended inference confidence: 0.40 (CLI --conf 0.4). 0.50 is a close alternative.

Code: regnetx-det (standalone sibling of dfine-det and convnextv2-det).

This is a domain fine-tune on Hebrew/Samaritan manuscripts from the native RegNetX Stage-0 foundation. For the multiscript Stage-0 checkpoint, see johnlockejrr/regnetx-8gf-polyline-baseline-stage0 (cbad_f1_max ≈ 0.936 on the shared Stage-0 val split; 0.953 zero-shot on this Hebrew+Samaritan val).

Model summary

Detection is set prediction of scored polylines. Polygonal line environments (BLLA) are applied only at PAGE/ALTO serialize time.

Intended use

Use for

• Hebrew square-script and Samaritan manuscript baseline detection
• PAGE-XML / ALTO export for HTR pipelines (RTL: --text-direction horizontal-rl)
• Research comparing RegNetX vs ConvNeXt / D-FINE polyline siblings on the same Arrow

Not for

• Official ICDAR cBAD 2019 test claims (this val split is Hebrew+Samaritan only)
• Production OCR transcription (lines only; no text recognition)
• Layout region detection (paragraphs, tables, illustrations)
• Treating cbad_f1 @ conf 0.1 as the quality number (see operating point below)

Training data

Same Arrow as the dfine-det / convnextv2-det Hebrew+Samaritan combined recipes:

Sources (PAGE-XML baselines):

• Hebrew medieval manuscripts (Hebrew_Medieval-seg)
• Samaritan manuscripts (sam_44_mss_pango_additional; manuscript-level split, seed 42)

Compiled with adaptive simplify → uniform cubic B-spline, (K=8), original-normalized coords (letterboxed at load time).

Training recipe

Config: configs/hebrew_samaritan_regnetx_stage0_finetune_balanced.yaml.

FPN-only recipes on the same data plateaued (~0.61 for 8GF, ~0.31 for 032). HybridEncoder + Stage-0 encoder transfer is the working path.

Training ran until early-stop (~epoch 90). The released snapshot is the monitored best (cbad_f1_max ≈ 0.969 at conf 0.40), not the last epoch.

Previous release used D-FINE Stage-0 warm-start (configs/hebrew_samaritan_stage0_hybrid.yaml), reaching cbad_f1_max ≈ 0.942. The native RegNetX Stage-0 balanced finetune improves this by +0.027 F1.

Evaluation (internal val)

Metric: cBAD-style line F1 — Hungarian matching of densified polylines with mean bidirectional Chamfer cost; TP if cost ≤ 20 px.

Released checkpoint

Conf sweep (locked-in test on best checkpoint)

cbad_f1 on the train progress bar is fixed conf=0.1. cbad_f1_max is the sweep max. Compare systems on cbad_f1_max (or F1 at a stated conf).

Stage-0 zero-shot baseline (same val, same metric)

Stage-0 best: conf=0.30 cbad_f1=0.953. Finetune gain: +0.016 F1 (0.953 → 0.969) at respective best confs.

Sibling bake-off (same Hebrew+Samaritan Arrow family)

Numbers are internal val, not a shared official test protocol.

How to use

Install

cd /path/to/regnetx-det
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Inference (PAGE XML)

huggingface-cli download johnlockejrr/regnetx-8gf-polyline-hebrew-samaritan \
  best_cbad_f1.safetensors --local-dir weights/hebrew_samaritan_hybrid

regnetx-det infer \
  -w weights/hebrew_samaritan_hybrid/best_cbad_f1.safetensors \
  --image page.jpg \
  -o page.xml \
  --format page \
  --conf 0.4

# ATR-style per-page (RTL Hebrew/Samaritan):
regnetx-det infer-page \
  -m weights/hebrew_samaritan_hybrid/best_cbad_f1.safetensors \
  -d cuda:0 \
  --conf 0.4 \
  --text-direction horizontal-rl \
  page.jpg

infer / infer-page default --conf is 0.5 (close to the sweep winner). Pass --conf 0.4 to match cbad_f1_max.

Fine-tune further

train:
  load: best_cbad_f1.safetensors   # this release (full weights)
  backbone: regnetx_080.tv2_in1k
  pretrained_backbone: false
  monitor: cbad_f1_max
  image_size: 1280

Files in this release

Limitations

• Val F1 is Hebrew+Samaritan holdout, not official cBAD 2019 test.
• (K=8) control points underfit strongly curved / torn lines.
• Export uses EMA when warmed during train; released file is the monitored best snapshot.
• HybridEncoder eval_spatial_size must match the canvas (1280). Do not run eval/infer at a different square size than training.

Citation & credits

Paper (prefer this)

Preprint: regnetx-det: Set-Prediction Polyline Baselines for Historical Documents with HybridEncoder Transfer (Zenodo, Aug 2026).

@misc{bors2026regnetxdet,
  author       = {Bors, Teodor},
  title        = {regnetx-det: Set-Prediction Polyline Baselines for Historical Documents with HybridEncoder Transfer},
  year         = {2026},
  month        = aug,
  publisher    = {Zenodo},
  version      = {v1},
  doi          = {10.5281/zenodo.22059333},
  url          = {https://doi.org/10.5281/zenodo.22059333}
}

D-FINE (HybridEncoder, deformable decoder, Stage-0 polyline weights)

@inproceedings{peng2025dfine,
  title     = {D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement},
  author    = {Peng, Yansong and Li, Hebei and Wu, Peixi and Zhang, Yueyi and Sun, Xiaoyan and Wu, Feng},
  booktitle = {The Thirteenth International Conference on Learning Representations},
  year      = {2025},
  url       = {https://arxiv.org/abs/2410.13842}
}

• Code: github.com/Peterande/D-FINE
• Stage-0 foundation: johnlockejrr/regnetx-8gf-polyline-baseline-stage0

RegNetX backbone

Radosavovic et al., Designing Network Design Spaces, CVPR 2020. Weights via timm regnetx_080.tv2_in1k.

This checkpoint (weights)

@software{regnetx_det_hebrew_samaritan_balanced,
  title  = {RegNetX-8GF HybridEncoder Polyline Baselines (Hebrew + Samaritan, Stage-0 Balanced Finetune)},
  author = {Bors, Teodor},
  year   = {2026},
  note   = {Fine-tuned from native RegNetX Stage-0; cbad_f1_max 0.969 at conf 0.40; see Zenodo preprint doi:10.5281/zenodo.22059333},
  url    = {https://huggingface.co/johnlockejrr/regnetx-8gf-polyline-hebrew-samaritan}
}

Additional notices

• Vendored D-FINE modules: THIRD_PARTY_NOTICES.md (Apache-2.0, © 2024 The D-FINE Authors).
• PAGE polygonalization: Kraken/BLLA subset (Apache-2.0, © Benjamin Kiessling) — export only, hardened with CLAHE/Sauvola fallback.
• Training images: respect each corpus license; this card does not redistribute the PAGE/JPEG data.

License

Apache License 2.0 for regnetx-det code and these weights. Downstream users must also comply with licenses of datasets used in further fine-tuning.