xberg-io/layout-models

xberg Layout Models

ONNX models used by xberg for document layout detection and table structure recognition.

Models

RT-DETR (Document Layout Detection)

Property	Value
Path	rtdetr/model.onnx
Size	169 MB
Precision	FP32
Architecture	RT-DETR v2 (Real-Time Detection Transformer)
Input	images: [batch, 3, 640, 640] f32 (ImageNet-normalized, letterboxed)
Input	orig_target_sizes: [batch, 2] i64 (original [height, width])
Outputs	labels i64, boxes f32 [batch, N, 4], scores f32
Classes	17 document layout classes
SHA256	3bf2fb0ee6df87435b7ae47f0f3930ec3dc97ec56fd824acc6d57bc7a6b89ef2

Layout Classes: Caption, Footnote, Formula, ListItem, PageFooter, PageHeader, Picture, SectionHeader, Table, Text, Title, DocumentIndex, Code, CheckboxSelected, CheckboxUnselected, Form, KeyValueRegion

PP-DocLayoutV3 (Document Layout Detection)

Property	Value
Path	pp_doclayout_v3/model.onnx
Size	130.5 MB (130,502,049 bytes)
Precision	FP32
Architecture	PP-DocLayoutV3, a DETR-style detector (PaddleDetection DETR) that also predicts per-detection masks and reading order
Input	image: [batch, 3, 800, 800] f32 (squash-resized to 800x800, divided by 255, no mean/std, CHW)
Input	im_shape: [batch, 2] f32 (resized height and width, [800, 800])
Input	scale_factor: [batch, 2] f32 ([800 / orig_h, 800 / orig_w])
Outputs	fetch_name_0 f32 [N, 7] (class_id, score, x1, y1, x2, y2, reading_order; boxes already in original pixel space), fetch_name_1 i32 [1] (row count), fetch_name_2 i32 [N, 200, 200] (masks)
ONNX	opset 17, IR 8, batch dimension dynamic, image size fixed at 800x800
Classes	25 document layout classes
SHA256	45bf71750b00739a41fc209f132eb104a4d6b5bb29483c9078164d8b87cf28ba

Layout Classes (index order): 0 abstract, 1 algorithm, 2 aside_text, 3 chart, 4 content, 5 display_formula, 6 doc_title, 7 figure_title, 8 footer, 9 footer_image, 10 footnote, 11 formula_number, 12 header, 13 header_image, 14 image, 15 inline_formula, 16 number, 17 paragraph_title, 18 reference, 19 reference_content, 20 seal, 21 table, 22 text, 23 vertical_text, 24 vision_footnote

pp_doclayout_v3/inference.yml is the upstream preprocessing and label configuration, unmodified.

TATR (Table Structure Recognition)

Property	Value
Path	tatr/model.onnx
Size	29 MB
Precision	INT8 quantized
Architecture	DETR (DEtection TRansformer) — non-autoregressive object detection
Input	pixel_values: [batch, 3, H, W] f32 (variable size, typically 800×800)
Outputs	logits f32 [batch, 125, 7] (class probabilities), pred_boxes f32 [batch, 125, 4] (normalized cx/cy/w/h)
Classes	7 classes (see below)
SHA256	c11f4033da75e9c4d41c403ef356e89caa0a37a7d111b55461e7d5ba856bb6b6

Table Structure Classes: 0. table — entire table region

1. table column — column span
2. table row — row span
3. table column header — header row cells
4. table projected row header — projected row header
5. table spanning cell — cells spanning multiple rows/columns
6. no object — background

Attribution & Provenance

RT-DETR

This model is mirrored from docling-project/docling-layout-heron-onnx, created by the Docling team at IBM Research.

• Original repository: docling-project/docling-layout-heron-onnx
• License: Apache-2.0
• Architecture paper: Zhao et al., "DETRs Beat YOLOs on Real-time Object Detection" (arXiv:2304.08069)
• Training data: DocLayNet and internal IBM document datasets

TATR (Table Transformer)

This model is based on microsoft/table-transformer-structure-recognition by Microsoft Research. The ONNX conversion was produced by Xenova/table-transformer-structure-recognition using HuggingFace Optimum. Quantized to INT8 for inference efficiency.

• Original repository: microsoft/table-transformer-structure-recognition
• ONNX source: Xenova/table-transformer-structure-recognition
• License: MIT
• Architecture paper: Smock et al., "PubTables-1M: Towards comprehensive table extraction from unstructured documents" (arXiv:2110.00061)
• Training data: PubTables-1M dataset
• Quantization: INT8 (dynamic quantization via ONNX Runtime)

PP-DocLayoutV3

pp_doclayout_v3/model.onnx and pp_doclayout_v3/inference.yml are byte-for-byte copies of inference.onnx and inference.yml from the official ONNX release published by the PaddlePaddle team. They are re-hosted here unmodified so that a download is pinned to a Hub revision of this repository.

• ONNX source: PaddlePaddle/PP-DocLayoutV3_onnx at revision 46bbdf188bb0a772c08aed74882ce7e51a8f1ea6
• Original weights: PaddlePaddle/PP-DocLayoutV3
• Project: PaddlePaddle/PaddleOCR (PP-DocLayoutV3 is the layout analysis module of PaddleOCR-VL-1.5)
• License: Apache-2.0
• Technical report: Cui et al., "PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing" (arXiv:2601.21957)
• Conversion: performed by PaddlePaddle; no conversion or modification by xberg. Copyright PaddlePaddle Authors.

Usage

These models are automatically downloaded and cached by the xberg document extraction library. See the documentation for details.

License

• RT-DETR: Apache-2.0 License
• TATR: MIT License
• PP-DocLayoutV3: Apache-2.0 License