ocr-trains-dataset.py

import datasets
import pandas as pd

_CITATION = """\
@InProceedings{huggingface:dataset,
title = {ocr-trains-dataset},
author = {TrainingDataPro},
year = {2023}
}
"""

_DESCRIPTION = """\
The dataset is a collection of images along with corresponding bounding box annotations
that are specifically curated for **detecting pigs' heads** in images. The dataset
covers different *pig breeds, sizes, and orientations*, providing a comprehensive
representation of pig appearances.
The pig detection dataset provides a valuable resource for researchers working on pig
detection tasks. It offers a diverse collection of annotated images, allowing for
comprehensive algorithm development, evaluation, and benchmarking, ultimately aiding in
the development of accurate and robust models.
"""
_NAME = "ocr-trains-dataset"

_HOMEPAGE = f"https://huggingface.co/datasets/TrainingDataPro/{_NAME}"

_LICENSE = ""

_DATA = f"https://huggingface.co/datasets/TrainingDataPro/{_NAME}/resolve/main/data/"


class OcrTrainsDataset(datasets.GeneratorBasedBuilder):
    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "id": datasets.Value("int32"),
                    "image": datasets.Image(),
                    "bboxes": datasets.Value("string"),
                }
            ),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        images = dl_manager.download(f"{_DATA}images.tar.gz")
        annotations = dl_manager.download(f"{_DATA}{_NAME}.csv")
        images = dl_manager.iter_archive(images)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "images": images,
                    "annotations": annotations,
                },
            ),
        ]

    def _generate_examples(self, images, annotations):
        annotations_df = pd.read_csv(annotations)

        for idx, (image_path, image) in enumerate(images):
            yield idx, {
                "id": annotations_df.loc[annotations_df["image_name"] == image_path][
                    "image_id"
                ].values[0],
                "image": {"path": image_path, "bytes": image.read()},
                "bboxes": annotations_df["annotations"].iloc[idx],
            }