DocCornerDataset

A high-quality document corner detection dataset for training models to detect the four corners of documents in images. This dataset is optimized for building robust document scanning and perspective correction applications.

Dataset Examples

Training Set

Validation Set

Test Set

Green polygons show the annotated document corners

Dataset Description

This dataset contains images with document corner annotations, optimized for training robust document detection models. It uses the best-performing splits from an iterative dataset cleaning process with multiple quality validation steps.

Key Features

• High Quality Annotations: Labels refined through iterative cleaning with multiple teacher models
• Diverse Document Types: IDs, invoices, receipts, books, cards, and general documents
• Negative Samples: Includes images without documents for training robust classifiers
• No Overlap: Train, validation, and test splits are completely disjoint

Dataset Statistics

Split	Images	Description
train	32,968	Training set (cleaned iter3 + hard negatives)
validation	8,645	Validation set (cleaned iter3)
test	6,652	Held-out test set (no overlap with train/val)
Total	48,265

Data Sources and Licenses

This dataset is compiled from multiple open-source datasets. Please refer to the original dataset licenses before using this data.

MIDV Dataset (ID Cards)

Mobile Identity Document Video dataset for identity document detection and recognition.

Dataset	Images	License	Source
MIDV-500	~9,400	Research use	Website
MIDV-2019	~1,350	Research use	Website

Citation:

@article{arlazarov2019midv500,
  title={MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream},
  author={Arlazarov, V.V. and Bulatov, K. and Chernov, T. and Arlazarov, V.L.},
  journal={Computer Optics},
  volume={43},
  number={5},
  pages={818--824},
  year={2019}
}

@inproceedings{arlazarov2019midv2019,
  title={MIDV-2019: Challenges of the modern mobile-based document OCR},
  author={Arlazarov, V.V. and Bulatov, K. and Chernov, T. and Arlazarov, V.L.},
  booktitle={ICDAR},
  year={2019}
}

SmartDoc Dataset (Documents)

SmartDoc Challenge dataset for document image acquisition and quality assessment.

Dataset	Images	License	Source
SmartDoc	~1,380	Research use	Website

Citation:

@inproceedings{burie2015smartdoc,
  title={ICDAR 2015 Competition on Smartphone Document Capture and OCR (SmartDoc)},
  author={Burie, J.C. and Chazalon, J. and Coustaty, M. and others},
  booktitle={ICDAR},
  year={2015}
}

COCO Dataset (Negative Samples)

Common Objects in Context dataset used for negative samples (images without documents).

Dataset	Images	License	Source
COCO val2017	~4,300	CC BY 4.0	Website
COCO train2017	~11,400	CC BY 4.0	Website

Note: Excluded categories that could be confused with documents: book, laptop, tv, cell phone, keyboard, mouse, remote, clock.

Citation:

@inproceedings{lin2014coco,
  title={Microsoft COCO: Common Objects in Context},
  author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and others},
  booktitle={ECCV},
  year={2014}
}

Roboflow Universe (Various Documents)

Various document datasets from Roboflow Universe community.

Category	Datasets	License	Source
Documents	document_segmentation_v2, doc_scanner, doc_rida, documento	Various (check individual)	Roboflow Universe
Bills/Invoices	bill_segmentation, cs_invoice	Various (check individual)	Roboflow Universe
Receipts	receipt_detection, receipt_occam, receipts_coolstuff	Various (check individual)	Roboflow Universe
ID Cards	card_corner, card_4_class, id_card_skew, id_detections, idcard_jj	Various (check individual)	Roboflow Universe
Passports	segment_passport	Various (check individual)	Roboflow Universe
Books	book_reader, page_segmentation_tecgp, book_cmjt2	Various (check individual)	Roboflow Universe

Note: Roboflow datasets have various licenses. Please check the individual dataset pages on Roboflow Universe for specific license terms.

Features

Feature	Type	Description
image	Image	The document image (JPEG)
filename	string	Original filename for traceability
is_negative	bool	True if image contains no document
corner_tl_x	float32	Top-left corner X coordinate (normalized 0-1)
corner_tl_y	float32	Top-left corner Y coordinate (normalized 0-1)
corner_tr_x	float32	Top-right corner X coordinate (normalized 0-1)
corner_tr_y	float32	Top-right corner Y coordinate (normalized 0-1)
corner_br_x	float32	Bottom-right corner X coordinate (normalized 0-1)
corner_br_y	float32	Bottom-right corner Y coordinate (normalized 0-1)
corner_bl_x	float32	Bottom-left corner X coordinate (normalized 0-1)
corner_bl_y	float32	Bottom-left corner Y coordinate (normalized 0-1)

Corner Order

Corners are ordered clockwise starting from top-left:

1 (TL) -------- 2 (TR)
   |              |
   |   Document   |
   |              |
4 (BL) -------- 3 (BR)

Coordinate System

• Coordinates are normalized to the range [0, 1]
• To convert to pixel coordinates: pixel_x = corner_x * image_width
• Origin (0, 0) is at the top-left of the image

Negative Samples

Images with is_negative=True:

• Do not contain any document
• All corner coordinates are null
• Useful for training classifiers to reject non-document images

Usage

Loading the Dataset

from datasets import load_dataset

# Load all splits
dataset = load_dataset("mapo80/DocCornerDataset")

# Access specific splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

print(f"Train: {len(train_data)} samples")
print(f"Val: {len(val_data)} samples")
print(f"Test: {len(test_data)} samples")

Iterating Over Samples

for sample in dataset["train"]:
    image = sample["image"]  # PIL Image
    filename = sample["filename"]

    if not sample["is_negative"]:
        # Get corner coordinates (normalized 0-1)
        corners = [
            (sample["corner_tl_x"], sample["corner_tl_y"]),
            (sample["corner_tr_x"], sample["corner_tr_y"]),
            (sample["corner_br_x"], sample["corner_br_y"]),
            (sample["corner_bl_x"], sample["corner_bl_y"]),
        ]

        # Convert to pixel coordinates
        w, h = image.size
        corners_px = [(int(x * w), int(y * h)) for x, y in corners]

Visualizing Annotations

from PIL import Image, ImageDraw

def draw_corners(image, corners, color=(0, 255, 0), width=3):
    """Draw document corners on image."""
    draw = ImageDraw.Draw(image)
    w, h = image.size

    # Convert normalized to pixel coords
    points = [(int(c[0] * w), int(c[1] * h)) for c in corners]

    # Draw polygon
    for i in range(4):
        draw.line([points[i], points[(i+1) % 4]], fill=color, width=width)

    # Draw corner circles
    for p in points:
        r = 5
        draw.ellipse([p[0]-r, p[1]-r, p[0]+r, p[1]+r], fill=color)

    return image

# Example usage
sample = dataset["train"][0]
if not sample["is_negative"]:
    corners = [
        (sample["corner_tl_x"], sample["corner_tl_y"]),
        (sample["corner_tr_x"], sample["corner_tr_y"]),
        (sample["corner_br_x"], sample["corner_br_y"]),
        (sample["corner_bl_x"], sample["corner_bl_y"]),
    ]
    annotated = draw_corners(sample["image"].copy(), corners)
    annotated.show()

Training a Model (PyTorch Example)

import torch
from torch.utils.data import DataLoader
from datasets import load_dataset

dataset = load_dataset("mapo80/DocCornerDataset")

def collate_fn(batch):
    images = torch.stack([transform(s["image"]) for s in batch])

    # Stack corner coordinates (8 values per sample)
    corners = []
    for s in batch:
        if s["is_negative"]:
            corners.append(torch.zeros(8))
        else:
            corners.append(torch.tensor([
                s["corner_tl_x"], s["corner_tl_y"],
                s["corner_tr_x"], s["corner_tr_y"],
                s["corner_br_x"], s["corner_br_y"],
                s["corner_bl_x"], s["corner_bl_y"],
            ]))

    return images, torch.stack(corners)

train_loader = DataLoader(
    dataset["train"],
    batch_size=32,
    shuffle=True,
    collate_fn=collate_fn
)

Model Performance

Models trained on this dataset achieve the following performance:

Model	Input Size	mIoU (val)	mIoU (test)
MobileNetV2 (alpha=0.35)	224x224	0.9894	0.9826
MobileNetV2 (alpha=0.35)	256x256	0.9902	0.9819

mIoU = Mean Intersection over Union between predicted and ground truth quadrilaterals

Citation

If you use this dataset in your research, please cite this dataset and the original source datasets:

@dataset{doccornerdataset2025,
  author = {mapo80},
  title = {DocCornerDataset: Document Corner Detection Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/mapo80/DocCornerDataset}
}

Please also cite the original datasets used:

• MIDV-500/MIDV-2019 (Arlazarov et al., 2019)
• SmartDoc (Burie et al., 2015)
• COCO (Lin et al., 2014)

License

⚠️ This dataset is compiled from multiple sources with different licenses.

Source	License
MIDV-500/MIDV-2019	Research use only
SmartDoc	Research use only
COCO	CC BY 4.0
Roboflow datasets	Various (check individual datasets)

Before using this dataset, please review the licenses of the original datasets:

• MIDV-500
• MIDV-2019
• SmartDoc
• COCO
• Roboflow Universe (check individual datasets)

Acknowledgments

This dataset was created by combining and processing multiple open-source datasets. We thank the authors of MIDV, SmartDoc, COCO, and the Roboflow community for making their data available.

Related Projects

• DocCornerNet - Document corner detection model trained on this dataset