LibrePPDocLayoutV3l

PP-DocLayoutV3 (HGNetV2-L backbone, RT-DETR-style decoder with mask-enhanced queries and a global-pointer reading-order head) converted to the LibreYOLO checkpoint format. It detects 25 document layout classes (text, titles, tables, images, formulas, headers, footers, seals, ...) and ranks the regions in reading order.

from libreyolo import LibreYOLO

model = LibreYOLO("LibrePPDocLayoutV3l.pt")
r = model("page.png", conf=0.5)
for order, box, cls in zip(r.reading_order, r.boxes.xyxy, r.boxes.cls):
    print(int(order), r.names[int(cls)], box.tolist())

Source

Derived from PaddlePaddle/PP-DocLayoutV3_safetensors at revision e4f489dc536556fc1be02b973ba06756e0d4a2ac (model.safetensors), released by the PaddleOCR project. Copyright (c) PaddlePaddle Authors. Licensed under the Apache License 2.0.

Modifications

State-dict key prefix removal (model.) and LibreYOLO v1.0 metadata only. Learned parameters are unchanged. Class names follow the release's inference.yml label_list. See weights/convert_ppdoclayoutv3_weights.py in the LibreYOLO source repository. Raw head outputs (class logits, boxes, reading-order logits, masks) match the reference PyTorch implementation bit for bit on identical inputs.

This file: LibrePPDocLayoutV3l.pt, SHA-256 347f3f60e03a4ecc64184085709180b094df4c27345b3a97fd9ec1b4f78a1ef4.

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository.