BDRC/Tibetan-Modern-Book-Layout-Detection-DocLayout-YOLO

Tibetan Modern Book Layout Detection (DocLayout-YOLO)

A DocLayout-YOLO (OpenDataLab, YOLOv10-based) object detector that locates the four structural regions of a modern Tibetan book page — header, text-area, footnote, footer — as a preprocessing step for OCR and etext production.

• Training code, recipes & write-up: buda-base/tibetan-book-layout-analysis
• Dataset: BDRC/TDLA-Training-Dataset-v2 (gated, fair-use)

This is one of several architectures BDRC fine-tuned on the same labels and recipe to test how much the choice of architecture matters (see the blog post). The primary, production release is the RT-DETR-l fine-tune at BDRC/Tibetan-Modern-Book-Layout-Detection-RTDETR (MIT-licensed). This checkpoint is licensed AGPL-3.0, not MIT, because DocLayout-YOLO's own codebase (and the Ultralytics YOLO it builds on) is AGPL-3.0 — see License below before using it in a closed-source product.

Model description

This is a DocLayout-YOLO detector fine-tuned with the tam2col labelling scheme: text-area boxes merged into one envelope per page, except on genuine two-column pages, where it keeps one box per column. Header and footer are kept as separate classes (they can be combined losslessly downstream). Unlike the off-the-shelf DocStructBench checkpoint (which uses a 10-class generic document schema), this model was fine-tuned directly on BDRC's 4-class schema, so its output classes need no remapping.

Property	Value
Architecture	DocLayout-YOLO (YOLOv10-based, via doclayout_yolo)
Task	Object detection
Base checkpoint	DocStructBench pretrained (OpenDataLab)
Image size	1024 × 1024
Number of classes	4
Framework	doclayout_yolo (YOLOv10)
Weights file	doclayout_yolo_tibetan_book_layout.pt

Classes

ID	Class	Description
0	header	running title / marginal text at top or side
1	text-area	main body text (one box per column)
2	footnote	notes below the text area
3	footer	folio numbers / marginal text at bottom or side

Recommended usage — per-class confidence thresholds

Like the primary RT-DETR-l release, this detector is recall-happy, so the best operating point differs by class. These are each class's own max-F1 confidence from a native per-class sweep on the test set:

class	recommended conf
header (0)	0.42
text-area (1)	0.66
footnote (2)	0.27
footer (3)	0.48

If you need a single global threshold, 0.30 is the best compromise (it is also the operating point used for the cross-architecture comparison in the blog post).

Inference

from doclayout_yolo import YOLOv10

model = YOLOv10("doclayout_yolo_tibetan_book_layout.pt")
CLASS_CONF = {0: 0.42, 1: 0.66, 2: 0.27, 3: 0.48}  # header, text-area, footnote, footer
names = {0: "header", 1: "text-area", 2: "footnote", 3: "footer"}

results = model.predict("page.jpg", imgsz=1024, conf=min(CLASS_CONF.values()))
for r in results:
    for b, cf, cl in zip(r.boxes.xyxy.tolist(), r.boxes.conf.tolist(), r.boxes.cls.tolist()):
        cls = int(cl)
        if cf < CLASS_CONF[cls]:
            continue
        print(names[cls], round(cf, 3), b)

A ready-made CLI (infer.py) with the thresholds baked in is included in this repo.

Downloading the weights

from huggingface_hub import hf_hub_download
path = hf_hub_download("BDRC/Tibetan-Modern-Book-Layout-Detection-DocLayout-YOLO",
                       "doclayout_yolo_tibetan_book_layout.pt")

Performance

Evaluated on the held-out test split (860 images) of BDRC/TDLA-Training-Dataset-v2.

Native 4-class metrics

class	P	R	F1	mAP@0.5	mAP@0.5:0.95
header	0.959	0.948	0.954	0.965	0.716
text-area	0.983	0.981	0.982	0.995	0.976
footnote	0.930	0.889	0.909	0.929	0.725
footer	0.963	0.957	0.960	0.961	0.684
overall	—	—	0.951	0.963	0.775

The mAP columns are threshold-independent — they integrate over the full precision/recall curve (every confidence), so they do not depend on any operating threshold. The P / R / F1 columns are reported at each class's own max-F1 confidence (see the per-class thresholds above), not at a fixed threshold.

Canonical 3-class metrics

Header + footer are combined into one header-footer class (matched individually), text-area is compared as one merged envelope, and footnote is left as-is — a fair space in which every fine-tuned architecture in the blog post was compared.

class	best-F1
header-footer	0.948
text-area	0.996
footnote	0.897
mean F1	0.947 (@ conf 0.30)

This lands slightly below BDRC's primary RT-DETR-l fine-tune (mean F1 0.960), mainly on the footnote class.

Contamination (the metric that actually matters for OCR)

Of the ground-truth headers/footers and footnotes this model misses, the share that get folded into its predicted text-area box (silently corrupting downstream OCR) rather than dropped cleanly:

region	detected	folded into text-area
header/footer	96%	0.1%
footnote	87%	0%

For comparison, off-the-shelf systems in the same evaluation ranged from 1.2% to 56% on header/footer contamination alone — see the blog post for the full picture.

Training details

Parameter	Value
Base checkpoint	DocStructBench pretrained (OpenDataLab)
Image size	1024
Epochs	100 planned, early-stopped at epoch 61 (best @ epoch 41)
GPU	single NVIDIA A10G (24 GB)

• Dataset: BDRC/TDLA-Training-Dataset-v2 — 8,325 images (6,751 train / 714 val / 860 test), volume-level leakage-free splits, augmented images confined to train.
• Label variant (tam2col): text-area boxes merged per page except on two-column pages; built with data/build_curricula.py in the GitHub repo.

Intended use

Automatic layout detection of modern Tibetan book pages, as a preprocessing step for OCR pipelines, document digitization, structured text extraction, and digital-library indexing.

Limitations

• Trained on modern Tibetan books; performance on traditional pecha, manuscripts, or woodblock prints is not characterized and may be poor.
• Optimized for 1024 px input; very high-resolution scans may benefit from a higher imgsz.
• The footnote class is rare in the source material (≈1.4% of boxes); this is this checkpoint's weakest class (F1 0.909, mAP@0.5:0.95 0.725), somewhat below BDRC's primary RT-DETR-l fine-tune.
• Header/footer boxes are small and easy to over-predict — use the recommended per-class thresholds above.

License

The model weights are released under the GNU Affero General Public License v3.0 (AGPL-3.0), not the permissive MIT license used for BDRC's primary RT-DETR-l release. This is because DocLayout-YOLO's training code (and the Ultralytics YOLO codebase it is built on) is itself AGPL-3.0-licensed, and these weights are a derivative work produced by fine-tuning it. If you use this checkpoint (or a model derived from it) as part of a network service, the AGPL-3.0 requires you to make the complete corresponding source available to users of that service. If that doesn't work for your use case, use BDRC's MIT-licensed RT-DETR-l release or the Apache-2.0-licensed RF-DETR-L release instead — both score comparably on this benchmark. This is not legal advice; consult your own counsel for how AGPL-3.0 applies to your use case.

The page images used for training are not covered by any content license — they are BDRC library scans distributed on a fair-use basis. You are solely responsible for your own copyright / rights analysis before use; BDRC accepts no liability for misuse. See the dataset card for the full notice.

Acknowledgements

Developed by the Buddhist Digital Resource Center (BDRC) for the BDRC Etext Corpus, with annotations produced and consolidated on the Ultralytics platform.

Citation

@software{bdrc_tibetan_book_layout_doclayoutyolo_2026,
  title   = {Tibetan Modern Book Layout Detection (DocLayout-YOLO)},
  author  = {Buddhist Digital Resource Center (BDRC)},
  year    = {2026},
  url     = {https://huggingface.co/BDRC/Tibetan-Modern-Book-Layout-Detection-DocLayout-YOLO}
}