creative-graphic-design/layout-validity

1from typing import List, Union2 3import datasets as ds4import evaluate5import numpy as np6import numpy.typing as npt7from evaluate.utils.file_utils import add_start_docstrings8 9_DESCRIPTION = r"""\10Computes the ratio of valid elements to all elements in the layout, where the area within the canvas of a valid element must be greater than 0.1% of the canvas.11"""12 13_KWARGS_DESCRIPTION = """\14Args:15    predictions (`list` of `list` of `float`): A list of lists of floats representing normalized `ltrb`-format bounding boxes.16    gold_labels (`list` of `list` of `int`): A list of lists of integers representing class labels.17    canvas_width (`int`, *optional*): Width of the canvas in pixels. Can be provided at initialization or during computation.18    canvas_height (`int`, *optional*): Height of the canvas in pixels. Can be provided at initialization or during computation.19 20Returns:21    float: The ratio of valid elements to all elements (0.0 to 1.0). An element is considered valid if its area within the canvas is greater than 0.1% of the canvas area.22 23Examples:24    >>> import evaluate25    >>> import numpy as np26    >>> metric = evaluate.load("creative-graphic-design/layout-validity")27    >>> # Normalized bounding boxes (left, top, right, bottom)28    >>> predictions = [[[0.1, 0.1, 0.5, 0.5], [0.6, 0.6, 0.9, 0.9]]]29    >>> gold_labels = [[1, 2]]  # Non-zero labels indicate valid elements30    >>> result = metric.compute(predictions=predictions, gold_labels=gold_labels, canvas_width=512, canvas_height=512)31    >>> print(result)32    1.033"""34 35_CITATION = """\36@inproceedings{hsu2023posterlayout,37  title={Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout},38  author={Hsu, Hsiao Yuan and He, Xiangteng and Peng, Yuxin and Kong, Hao and Zhang, Qing},39  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},40  pages={6018--6026},41  year={2023}42}43"""44 45 46@add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)47class LayoutValidity(evaluate.Metric):48    def __init__(49        self,50        canvas_width: int | None = None,51        canvas_height: int | None = None,52        **kwargs,53    ) -> None:54        super().__init__(**kwargs)55        self.canvas_width = canvas_width56        self.canvas_height = canvas_height57 58    def _info(self) -> evaluate.EvaluationModuleInfo:59        return evaluate.MetricInfo(60            description=_DESCRIPTION,61            citation=_CITATION,62            inputs_description=_KWARGS_DESCRIPTION,63            features=ds.Features(64                {65                    "predictions": ds.Sequence(ds.Sequence(ds.Value("float64"))),66                    "gold_labels": ds.Sequence(ds.Sequence(ds.Value("int64"))),67                }68            ),69            codebase_urls=[70                "https://github.com/PKU-ICST-MIPL/PosterLayout-CVPR2023/blob/main/eval.py#L105-L127"71            ],72        )73 74    def _compute(75        self,76        *,77        predictions: Union[npt.NDArray[np.float64], List[List[float]]],78        gold_labels: Union[npt.NDArray[np.int64], List[int]],79        canvas_width: int | None = None,80        canvas_height: int | None = None,81    ) -> float:82        # パラメータの優先順位処理83        canvas_width = canvas_width if canvas_width is not None else self.canvas_width84        canvas_height = (85            canvas_height if canvas_height is not None else self.canvas_height86        )87 88        if canvas_width is None or canvas_height is None:89            raise ValueError(90                "canvas_width and canvas_height must be provided either "91                "at initialization or during computation"92            )93 94        predictions = np.array(predictions)95        gold_labels = np.array(gold_labels)96 97        predictions[:, :, ::2] *= canvas_width98        predictions[:, :, 1::2] *= canvas_height99 100        total_elements, empty_elements = 0, 0101 102        w = canvas_width / 100103        h = canvas_height / 100104 105        assert len(predictions) == len(gold_labels)106 107        for gold_label, prediction in zip(gold_labels, predictions):108            mask = (gold_label > 0).reshape(-1)109            mask_prediction = prediction[mask]110            total_elements += len(mask_prediction)111            for mp in mask_prediction:112                xl, yl, xr, yr = mp113                xl = max(0, xl)114                yl = max(0, yl)115                xr = min(canvas_width, xr)116                yr = min(canvas_height, yr)117 118                if abs((xr - xl) * (yr - yl)) < w * h * 10:119                    empty_elements += 1120 121        return 1 - empty_elements / total_elements122