Datasets#

doctr.datasets provides ready-to-use public datasets, synthetic data generators, and loaders for your own annotations. They all return samples in the format expected by the docTR models and training scripts.

Built-in datasets#

Public datasets are downloaded and extracted on first use (download=True) into ~/.cache/doctr/datasets.

from doctr.datasets import CORD

train_set = CORD(train=True, download=True)
sample = train_set[0]
image, target = sample.image, sample.target

The same dataset can be loaded for each task:

OCR (default)

Full pages, with the boxes and the transcriptions of the words:

train_set = CORD(train=True, download=True)
target = train_set[0].target  # {"boxes": array of shape (N, 4), "labels": [str, ...]}

Detection

Full pages, with the boxes of the words only:

train_set = CORD(train=True, download=True, detection_task=True)

Recognition

Word crops, with their transcription:

train_set = CORD(train=True, download=True, recognition_task=True)
image, word = train_set[0].image, train_set[0].target

Add use_polygons=True to get rotated boxes as (4, 2) polygons instead of straight boxes (for recognition, crops are then extracted along the rotated box).

Available datasets#

Dataset	Detection (train / test pages)	Recognition (train / test crops)	Notes
`FUNSD`	149 / 50	21,888 / 8,707	English forms
`SROIE`	626 / 360	33,608 / 19,342	Receipts, upper-case labels
`CORD`	800 / 100	19,370 / 2,186	Receipts
`IIIT5K`	2,000 / 3,000	2,000 / 3,000	Scene text
`SVT`	100 / 249	257 / 647	Street view, upper-case labels
`SVHN`	33,402 / 13,068	73,257 / 26,032	House numbers, digits only
`SynthText`	772,875 / 85,875	~7,100,000 / 707,470	Synthetic
`IC03`	246 / 249	1,156 / 1,107	Scene text
`IC13`	229 / 233	849 / 1,095	Scene text, manual download
`IMGUR5K`	7,149 / 796	207,901 / 22,672	Handwritten, manual download
`MJSynth`	—	7,581,382 / 1,337,891	Synthetic words, manual download
`IIITHWS`	—	7,141,797 / 793,533	Handwritten words, manual download
`WILDRECEIPT`	1,268 / 472	49,377 / 19,598	Receipts, manual download
`COCOTEXT`	13,880 / 3,261	59,130 / 13,258	Natural images, legible text only, manual download
`DocArtefacts`	2,700 / 300	—	Object detection: QR codes, bar codes, logos, photos

Datasets marked manual download must be fetched from their authors’ website first: pass the paths of the extracted files to the constructor (see the API reference).

Synthetic data#

Generators render random characters or words on the fly, which is useful to pre-train a model or to add an alphabet to it. Both are used by the reference training scripts when no dataset is given.

from doctr.datasets import VOCABS, CharacterGenerator, WordGenerator

# Single characters, e.g. to train a character classifier
chars = CharacterGenerator(vocab=VOCABS["german"], num_samples=1000)

# Random words of 1 to 12 characters
words = WordGenerator(vocab=VOCABS["german"], min_chars=1, max_chars=12, num_samples=10000)
sample = words[0]  # sample.image: rendered word, sample.target: its text

Both accept a font_family (one font or a list of fonts) to render with your own fonts.

Your own data#

Load annotated data stored in the training format of each task:

Class	Task	Expected files
`DetectionDataset`	Text detection	`images/` + `labels.json` with polygons (optionally per class)
`RecognitionDataset`	Text recognition	word images + `labels.json` mapping each image to its text
`OCRDataset`	End-to-end OCR	`images/` + `labels.json` with boxes and transcriptions
`LayoutDataset`	Layout analysis	`images/` + `labels.json` with polygons and classes
`TableStructureDataset`	Table structure	`images/` + `labels.json` with cells and their logical coordinates
`OrientationDataset`	Orientation classification	a folder of upright images

from doctr.datasets import DetectionDataset, OCRDataset, RecognitionDataset

det_set = DetectionDataset(img_folder="path/to/images", label_path="path/to/labels.json")
reco_set = RecognitionDataset(img_folder="path/to/images", labels_path="path/to/labels.json")
ocr_set = OCRDataset(img_folder="path/to/images", label_file="path/to/labels.json")

Transforms and batching#

Pass transformations with img_transforms (applied to the image) and sample_transforms (applied to the image and its target, e.g. rotations). Datasets return images of different sizes, so resize them before batching. Batch with a PyTorch DataLoader and the collate_fn of the dataset:

from torch.utils.data import DataLoader

from doctr import transforms as T
from doctr.datasets import RecognitionDataset

train_set = RecognitionDataset(
    img_folder="path/to/images",
    labels_path="path/to/labels.json",
    img_transforms=T.Resize((32, 128), preserve_aspect_ratio=True),
)

train_loader = DataLoader(train_set, batch_size=32, shuffle=True, collate_fn=train_set.collate_fn)
images, targets = next(iter(train_loader))  # images: (32, 3, 32, 128) tensor, targets: list of 32 strings

doctr.transforms provides document-specific augmentations (shadows, glare, perspective, rotations…). See the API reference and the reference training scripts for complete pipelines.