Add as_recognition_dataset to VisionDataset

[felixdittrich92]

🚀 The feature

Add a method to the VisionDataset which make it possible to use each dataset also for recognition task.
Todo would be crop each box as image with there plain label and all exsisting chars as vocab.
A must have would be to merge multible datasets
@fg-mindee
wdyt ?

Motivation, pitch

Provide datasets also for recognition would be a big benefit for training part and i think a very good way to validate the results

Alternatives

No response

Additional context

No response

[fg-mindee]

Hi @felixdittrich92!

Yup, I agree but I think the best way to achieve this would be to have a constructor flag for each dataset:

• all dataset don't have the same task annotations
• so when the user request text detection annotations from a text recognition one, we'll have to throw an error
• in compatible cases, I would say that having a class attribute specifying the possible compatible tasks. And the constructor would take a target_task that would default to the one of the original dataset. And if it's different we perform the conversion

What do you think?

[felixdittrich92]

@fg-mindee
sounds good to me but can we add an minimalistic example at this issue ?
🤗

[fg-mindee]

I was thinking about something quite simple:

from doctr.datasets import FutureOCRDataset

ocr_ds = FutureOCRDataset("/path/to/root")
detection_ds = FutureOCRDataset("/path/to/root", target_task="text-detection")
recognition_ds = FutureOCRDataset("/path/to/root", target_task="text-recognition")

[felixdittrich92]

class FutureOCRDataset(VisionDataset):
    """FutureOCRDataset dataset from
    `"XYZ"
    <https://XYZ>`_.

    Example::
        >>> from doctr.datasets import FutureOCRDataset
        >>> train_set = FutureOCRDataset(train=True, download=True, target_task="text-recognition")
        >>> img, target = train_set[0]

    Args:
        train: whether the subset should be the training one
        sample_transforms: composable transformations that will be applied to each image
        rotated_bbox: whether polygons should be considered as rotated bounding box (instead of straight ones)
        target_task: the task for which the dataset is prepared
        **kwargs: keyword arguments from `VisionDataset`.
    """

    URL = 'XYZ'
    SHA256 = '123'

    def __init__(
        self,
        train: bool = True,
        target_task: str = "text-detection",
        sample_transforms: Optional[Callable[[Any], Any]] = None,
        rotated_bbox: bool = False,
        **kwargs: Any,
    ) -> None:

        super().__init__(url=self.URL, file_name='FutureOCRDataset', file_hash=self.SHA256, extract_archive=True, **kwargs)
        self.sample_transforms = sample_transforms
        self.train = train

        self.data: List[Tuple[Path, Dict[str, Any]]] = []
        
        if target_task == "text-recognition":
            
           self.data.append(_raw_path, dict(labels=_raw_label))
        else:
           self.data.append((_raw_path, dict(boxes=np.asarray(box_targets, dtype=np_dtype), labels=tuple(_raw_label))))

        OR
        if target_task == "text-recognition":
           raise ValueError("XYZ")

    def extra_repr(self) -> str:
        return f"train={self.train}"

   def prepare_for_recognition(self, data) -> Tuple[Path, str]:
        pass