D4LA — Diverse Document Layout Analysis

D4LA is a large-scale, diverse, and manually annotated benchmark dataset for document layout analysis (DLA), introduced at ICCV 2023 alongside the Vision Grid Transformer (VGT) model. It is sourced from the real-world IIT-CDIP document collection and covers 12 document types with 27 fine-grained layout categories.

D4LA is considered the most diverse and richly annotated DLA benchmark, designed to reflect the complexity of real-world document layouts across domains.

---

Splits

Split	Images	Annotation Instances
train	8,868	~84K
test	2,224	~21K
Total	11,092	~105K

---

Dataset Features

Each example in this dataset contains the following fields:

Column	Type	Description
image	Image	Document page image — viewable directly in the HF browser
file_name	string	Original filename of the image
category_ids	Sequence[int]	Class ID for each annotated instance (0–26)
category_names	Sequence[string]	Human-readable class name for each instance
quads	Sequence[Sequence[float]]	Quadrilateral coordinates per instance: [x1, y1, x2, y2, x3, y3, x4, y4], normalized to [0, 1]

Note: Coordinates are normalized — multiply by image width/height to get pixel values.

---

27 Layout Categories

ID	Name	Description
0	DocTitle	The main title at the head of the document, typically bold or underlined
1	ListText	A paragraph containing special enumeration or bullet symbols
2	LetterHead	The inside address at the beginning of a letter
3	Question	Questions (e.g., true/false items in questionnaires)
4	RegionList	Wireless form list regions
5	TableName	The caption or label of a table
6	FigureName	The caption or label of a figure
7	Footer	The footer region of a document body
8	Number	A special vertical number that is not document content
9	ParaTitle	The title of a paragraph section (similar to "Section" in DocBank)
10	RegionTitle	The title of a broader region block
11	LetterDear	The salutation/greeting line in a letter
12	OtherText	Uncategorized or miscellaneous text
13	Abstract	The abstract section of a document
14	Table	A table region
15	Equation	A formula or equation, often including equation numbers
16	PageHeader	The header area at the top of a page
17	Catalog	A table of contents block (text + page numbers)
18	ParaText	A general multi-line paragraph
19	Date	A date string
20	LetterSign	The complimentary close and signature at the end of a letter
21	RegionKV	A key-value pair region
22	Author	Author name(s)
23	Figure	A figure or image region
24	Reference	A bibliography or reference block
25	PageFooter	The footer area at the bottom of a page
26	PageNumber	A page number, typically in the header or footer

---

12 Document Types

D4LA covers the following document categories from IIT-CDIP:

Letter, Memo, Email, Scientific Report, Scientific Publication, Specification, File Folder, Form, Handwritten, Advertisement, Budget, Invoice

---

Usage

from datasets import load_dataset

ds = load_dataset("tuandunghcmut/D4LA")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['image', 'file_name', 'category_ids', 'category_names', 'quads'], num_rows: 8868})
#     test:  Dataset({features: ['image', 'file_name', 'category_ids', 'category_names', 'quads'], num_rows: 2224})
# })

# Access a sample
sample = ds["train"][0]
img = sample["image"]             # PIL Image
print(sample["category_names"])   # e.g. ['ParaText', 'DocTitle', 'Table']
print(sample["quads"])            # e.g. [[x1,y1,x2,y2,x3,y3,x4,y4], ...]

Convert quads → axis-aligned bounding boxes

def quad_to_bbox(quad, width, height):
    """Convert normalized quad to pixel [xmin, ymin, xmax, ymax]."""
    xs = [quad[i] * width  for i in range(0, 8, 2)]
    ys = [quad[i] * height for i in range(1, 8, 2)]
    return [min(xs), min(ys), max(xs), max(ys)]

for quad, name in zip(sample["quads"], sample["category_names"]):
    bbox = quad_to_bbox(quad, img.width, img.height)
    print(f"{name}: {bbox}")

Visualize annotations

from PIL import ImageDraw

draw = ImageDraw.Draw(img)
for quad, name in zip(sample["quads"], sample["category_names"]):
    pts = [(quad[i] * img.width, quad[i+1] * img.height) for i in range(0, 8, 2)]
    draw.polygon(pts, outline="red")
    draw.text(pts[0], name, fill="red")
img.show()

---

Annotation Format Details

Labels use a YOLO-style quadrilateral format (not the standard axis-aligned box):

class_id  x1 y1  x2 y2  x3 y3  x4 y4

• All coordinates are normalized to [0.0, 1.0] relative to image width/height
• Points are ordered: top-left → top-right → bottom-right → bottom-left (clockwise)
• Each image may have zero or more annotation instances

---

License

This dataset is distributed under CDLA-Permissive-2.0. Please refer to the original dataset repository for full terms of use.

---

Citation

@InProceedings{Da_2023_ICCV,
    author    = {Da, Cheng and Wang, Chuwei and Zheng, Qi and Li, Chunhui},
    title     = {Vision Grid Transformer for Document Layout Analysis},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2023},
    pages     = {19530--19540}
}

---

Links

• ICCV 2023 Paper
• Papers with Code
• Supplementary Materials