dfine-det-large-baseline-stage0

Stage-0 multiscript pretrained weights for dfine-det — a D-FINE Large (HGNetv2-B4) detector adapted to predict text-line baselines as polylines (B-spline control points), not boxes.

Hub repo: johnlockejrr/dfine-det-large-baseline-stage0
Primary file: best_cbad_f1.safetensors (~116 MB)

This checkpoint is intended as a generic baseline pretrain for downstream Stage-1 fine-tunes (e.g. Hebrew/Samaritan, ICDAR cBAD, other PAGE/ALTO corpora).

---

Model summary

Architecture	PolylineDFINE — D-FINE detection core + polyline head
Size class	Large (D-FINE-L / backbone B4)
Parameters	~30.1 M trainable
Queries	300
Geometry	(K=8) cubic B-spline control points per line + height
Canvas	(1280\times1280) letterbox
Init	Official Peterande/D-FINE dfine_l_obj2coco_e25.pth (strict=False)
Task	Document baseline / text-line detection → PAGE or ALTO export

Detection is set prediction of scored polylines. Polygonal line environments (BLLA) are applied only at PAGE/ALTO serialize time, not in the training objective.

---

Intended use

Use for

• Starting point for Stage-1 fine-tuning on a target script/corpus
• Zero-shot / few-shot probing of baselines on heterogeneous manuscripts (expect domain gap)
• Research on polyline baseline detectors built on D-FINE

Not for

• Claiming official ICDAR cBAD 2019 test numbers from these Stage-0 val scores (see below)
• Production OCR transcription (this model detects lines; it does not recognize text)
• Replacing layout region detectors (paragraphs, tables, etc.)

---

Training data (Stage-0)

Multiscript PAGE-XML / ALTO baselines under baseline_pretrain_raw (HTR-United–style open corpora + private packs), compiled to Arrow with adaptive simplify → B-spline (K=8).

Split	Pages
Train	46 513
Val (holdout)	1 923 (~5%, seed 42)
Total XMLs inventoried	48 436 across 29 corpora

Corpora (train / val pages)

Corpus	Train	Val	Scripts / notes
de-dh-unibe	14309	80	German manuscripts
ar-Bulac	7461	394	Arabic
he-Sofer_Mahir	5474	289	Hebrew
se-Riksarkivet_page	3861	217	Swedish archival
cu-Generic_Church_Slavonic_Handwriting	2638	80	Church Slavonic
smp-private	2143	113	Samaritan (private)
la-cmmhwr26	2052	275	Latin
eu-cBAD-ICDAR2017	1859	98	cBAD 2017
eu-cBAD-ICDAR2019	1398	74	cBAD 2019 (subset in pretrain mix)
syr-MIDRASH	1055	56	Syriac
ar-Muharaf	907	48	Arabic
ar-OpenITI-arabic_print_data	819	45	Arabic print (ALTO)
fi-FCR_500	475	25	Finnish
de-ICFHR-2016	427	21	German (list entry; image availability varies)
sa-…Pracalit…Sanskrit_and_Newar…	417	22	Sanskrit / Newar
syr-HTR_Winter_School_2024	259	14	Syriac
la-HTR_Winter_School_Vienna	255	16	Latin
he-BiblIA	125	7	Hebrew
la-CREMMA	100	21	Latin
+ smaller packs	…	…	Greek, Czech, Polish, Judeo-Spanish, Icelandic, Estonian, papyri, etc.

Important: ICDAR cBAD pages appear inside this pretrain mixture. Stage-0 validation F1 is an internal multiscript holdout, not a fair official cBAD 2019 test score. For apples-to-apples cBAD bake-offs, use a separate Stage-1-cBAD fine-tune and the official test split.

---

Training recipe

Hyperparameter	Value
Init weights	dfine-l-obj2coco (Peterande/D-FINE)
Optimizer	AdamW; base LR (1\times10^{-4}) (backbone (0.1\times))
Schedule	Linear warmup + cosine
Precision	bf16-mixed
Effective batch	32 (droplet: micro-batch 32 × accum 1 on MI300X-class)
Epochs (cap)	80; early-stop on cbad_f1_max (patience 12, min 15)
Augment	Mild photometric + ±2° rotation (OOB lines dropped)
Match distance	20 px on the 1280 canvas
Monitor	cbad_f1_max over conf ∈ {0.1…0.6}
Seed	42

Config references: configs/baseline_pretrain.yaml, configs/baseline_pretrain_droplet.yaml.

Training was stopped around epoch 39 (credits / wall-clock), after the monitor had plateaued near 0.89.

---

Evaluation (Stage-0 val)

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

Released checkpoint

Metric	Value	Notes
cbad_f1_max	≈ 0.893	Best monitor; operating conf typically 0.5
Weights file	best_cbad_f1.safetensors	Also tagged best_0.8930.safetensors
Source	checkpoints/best-epoch=39.ckpt	Converted with dfine-det convert

Curves / logs (bundled)

• history.csv — TensorBoard export for epochs 1–38 (peak in CSV: 0.8896 @ ep38, conf 0.5)
• stage0_metrics.png / .pdf — training curves from that CSV
• At ep38 @ conf 0.5: precision 0.943, recall 0.842, mean Chamfer ≈ 4.32 px

Fixed conf 0.1 F1 is substantially lower than cbad_f1_max — always sweep (or use ~0.5) before comparing systems.

---

How to use

Install

pip install -e ".[dev]"   # from the dfine-det repo; see QUICKSTART for torch/CUDA notes

Inference (PAGE XML)

dfine-det -d cuda:0 infer \
  --weights best_cbad_f1.safetensors \
  --image page.jpg \
  -o page.xml \
  --format page \
  --conf 0.5

Fine-tune (Stage-1)

train:
  load: best_cbad_f1.safetensors   # this Stage-0 Large pretrain
  backbone: B4
  lr: 5.0e-5
  loss_height: 3.0
  cost_y: 3.0
  monitor: cbad_f1_max

dfine-det --config your_stage1.yaml -d cuda:0 train --augment

---

Files in this release

File	Description
best_cbad_f1.safetensors	Recommended Stage-0 Large weights
best_0.8930.safetensors	Same snapshot, score-tagged filename
history.csv	Per-epoch val metrics (epochs 1–38 export)
stage0_metrics.png	Summary figure
README.md	This model card

---

Limitations

• Val F1 is not official cBAD 2019 test performance.
• Pretrain includes cBAD pages → do not treat Stage-0 zero-shot cBAD scores as an external bake-off.
• (K=8) control points underfit strongly curved / torn lines; consider higher (K) in future Stage-0 runs.
• EMA is used for export in training runs; released safetensors are the monitored best snapshot.
• Some listed corpora historically had missing sidecars (e.g. de-ICFHR-2016); compile filtered what resolved on disk.

---

Citation & credits

Please credit D-FINE (architecture & COCO/Objects365 pretrain)

This work builds on official D-FINE by Yansong Peng et al. (ICLR 2025 Spotlight). Backbone, hybrid encoder, deformable attention, and CDN utilities are adapted from their Apache-2.0 codebase; detection init uses their D-FINE-L Objects365→COCO checkpoint.

@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
• Weights: huggingface.co/Peterande/D-FINE (dfine_l_obj2coco_e25.pth)

This Stage-0 polyline adaptation (dfine-det)

@software{dfine_det_stage0_large,
  title   = {dfine-det Large Stage-0: Multiscript Polyline Baseline Pretrain},
  author  = {John Locke Jrr, johnlockejrr},
  year    = {2026},
  note    = {Initialized from Peterande D-FINE-L; polyline baseline head and PAGE/ALTO training pipeline},
  url     = {https://huggingface.co/johnlockejrr/dfine-det-large-baseline-stage0}
}

Acknowledgements

We thank AMD for complimentary GPU compute on the AMD Developer Cloud (AMD Instinct™ MI300X), which supported Stage-0 pretraining of this model.

Additional notices

• Vendored D-FINE modules: see THIRD_PARTY_NOTICES.md (Apache-2.0, © 2024 The D-FINE Authors).
• PAGE polygonalization helpers derive from Kraken/BLLA (Apache-2.0, © Benjamin Kiessling) — used at export, not for this pretrain objective.
• Training corpora: respect each dataset’s original license (HTR-United / Zenodo / institutional terms). Private packs (smp-private, et-private, …) are not redistributed with this card; only model weights trained with them are released if your license allows.

---

License

Apache License 2.0 for the dfine-det code and these weights, consistent with D-FINE’s Apache-2.0 release. Downstream users must also comply with licenses of any datasets used in further fine-tuning.