Legacy Metrics

Starting with 0.23.0, a new metrics module is being introduced to supervision. Metrics here are part of the legacy evaluation API and will be deprecated in the future.

Install the metrics extra before using this page's APIs:

pip install "supervision[metrics]"

ConfusionMatrix

supervision.metrics.detection.ConfusionMatrix dataclass

Confusion matrix for object detection tasks.

Attributes:

Name	Type	Description
matrix	NDArray[int32]	An 2D np.ndarray of shape (len(classes) + 1, len(classes) + 1) containing the number of TP, FP, FN and TN for each class.
classes	list[str]	Model class names.
conf_threshold	float	Detection confidence threshold between 0 and 1. Detections with lower confidence will be excluded from the matrix.
iou_threshold	float	Detection IoU threshold between 0 and 1. Detections with lower IoU will be classified as FP.
metric_target	MetricTarget	The type of detection data used for IoU computation. Informational metadata set by from_detections and from_tensors. Excluded from __eq__ comparisons — two ConfusionMatrix instances with identical matrix, classes, conf_threshold, and iou_threshold compare as equal regardless of metric_target.

Source code in src/supervision/metrics/detection.py

@dataclass
class ConfusionMatrix:
    """
    Confusion matrix for object detection tasks.

    Attributes:
        matrix: An 2D `np.ndarray` of shape `(len(classes) + 1, len(classes) + 1)`
            containing the number of `TP`, `FP`, `FN` and `TN` for each class.
        classes: Model class names.
        conf_threshold: Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded from the matrix.
        iou_threshold: Detection IoU threshold between `0` and `1`.
            Detections with lower IoU will be classified as `FP`.
        metric_target: The type of detection data used for IoU computation.
            Informational metadata set by `from_detections` and `from_tensors`.
            Excluded from `__eq__` comparisons — two `ConfusionMatrix` instances
            with identical `matrix`, `classes`, `conf_threshold`, and
            `iou_threshold` compare as equal regardless of `metric_target`.
    """

    matrix: npt.NDArray[np.int32]
    classes: list[str]
    conf_threshold: float
    iou_threshold: float
    metric_target: MetricTarget = MetricTarget.BOXES

    def __eq__(self, other: object) -> bool:
        if not isinstance(other, ConfusionMatrix):
            return NotImplemented
        return (
            np.array_equal(self.matrix, other.matrix)
            and self.classes == other.classes
            and self.conf_threshold == other.conf_threshold
            and self.iou_threshold == other.iou_threshold
        )

    __hash__ = None  # type: ignore[assignment]

    @classmethod
    def from_detections(
        cls,
        predictions: list[Detections],
        targets: list[Detections],
        classes: list[str],
        conf_threshold: float = 0.3,
        iou_threshold: float = 0.5,
        metric_target: MetricTarget = MetricTarget.BOXES,
    ) -> ConfusionMatrix:
        """
        Calculate confusion matrix based on predicted and ground-truth detections.

        Args:
            targets: Detections objects from ground-truth.
            predictions: Detections objects predicted by the model.
            classes: Model class names.
            conf_threshold: Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold: Detection IoU threshold between `0` and `1`.
                Detections with lower IoU will be classified as `FP`.
            metric_target: The type of detection data to use.
                Supports `MetricTarget.BOXES` (default) and
                `MetricTarget.ORIENTED_BOUNDING_BOXES`. When using
                `MetricTarget.ORIENTED_BOUNDING_BOXES`, each `Detections`
                object must include OBB coordinates in
                `detections.data[ORIENTED_BOX_COORDINATES]` as a float32
                array of shape `(N, 8)` (flat) or `(N, 4, 2)` (as stored by
                `from_ultralytics`); both are normalised to `(N, 8)` internally.
                `MetricTarget.MASKS` is not supported.

        Returns:
            New instance of ConfusionMatrix.

        Examples:
            ```pycon
            >>> import numpy as np
            >>> import supervision as sv
            >>> targets = [
            ...     sv.Detections(
            ...         xyxy=np.array([[0, 0, 10, 10], [50, 50, 60, 60]]),
            ...         class_id=np.array([0, 0])
            ...     )
            ... ]
            >>> predictions = [
            ...     sv.Detections(
            ...         xyxy=np.array([[0, 0, 10, 10], [100, 100, 110, 110]]),
            ...         class_id=np.array([0, 0]),
            ...         confidence=np.array([0.9, 0.8])
            ...     )
            ... ]
            >>> confusion_matrix = sv.ConfusionMatrix.from_detections(
            ...     predictions=predictions,
            ...     targets=targets,
            ...     classes=['person']
            ... )
            >>> confusion_matrix.matrix
            array([[1, 1],
                   [1, 0]], dtype=int32)

            ```
        """
        prediction_tensors = []
        target_tensors = []
        for prediction, target in zip(predictions, targets):
            prediction_tensors.append(
                detections_to_tensor(
                    prediction, with_confidence=True, metric_target=metric_target
                )
            )
            target_tensors.append(
                detections_to_tensor(
                    target, with_confidence=False, metric_target=metric_target
                )
            )
        return cls.from_tensors(
            predictions=prediction_tensors,
            targets=target_tensors,
            classes=classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
            metric_target=metric_target,
        )

    @classmethod
    def from_tensors(
        cls,
        predictions: list[npt.NDArray[np.float32]],
        targets: list[npt.NDArray[np.float32]],
        classes: list[str],
        conf_threshold: float = 0.3,
        iou_threshold: float = 0.5,
        metric_target: MetricTarget = MetricTarget.BOXES,
    ) -> ConfusionMatrix:
        """
        Calculate confusion matrix based on predicted and ground-truth detections.

        Args:
            predictions: Each element of the list describes a single
                image and has `shape = (M, 6)` or `shape = (M, 10)` depending on
                `metric_target`.
                If `MetricTarget.BOXES`, each row is in
                `(x_min, y_min, x_max, y_max, class, conf)` format.
                If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
                `(x1, y1, x2, y2, x3, y3, x4, y4, class, conf)` format.
            targets: Each element of the list describes a single
                image and has `shape = (N, 5)` or `shape = (N, 9)` depending on
                `metric_target`.
                If `MetricTarget.BOXES`, each row is in
                `(x_min, y_min, x_max, y_max, class)` format.
                If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
                `(x1, y1, x2, y2, x3, y3, x4, y4, class)` format.
            classes: Model class names.
            conf_threshold: Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold: Detection iou threshold between `0` and `1`.
                Detections with lower iou will be classified as `FP`.
            metric_target: The type of detection data to use.
                Determines expected tensor shapes (see Args above for column
                layouts). `MetricTarget.MASKS` is not supported.

        Returns:
            New instance of ConfusionMatrix.

        Examples:
            ```pycon
            >>> import supervision as sv
            >>> import numpy as np
            >>> targets = [
            ...     np.array([
            ...         [0.0, 0.0, 3.0, 3.0, 0],
            ...         [2.0, 2.0, 5.0, 5.0, 0],
            ...         [6.0, 1.0, 8.0, 3.0, 1],
            ...     ])
            ... ]
            >>> predictions = [
            ...     np.array([
            ...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
            ...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
            ...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
            ...     ])
            ... ]
            >>> confusion_matrix = sv.ConfusionMatrix.from_tensors(
            ...     predictions=predictions,
            ...     targets=targets,
            ...     classes=['person', 'dog']
            ... )
            >>> confusion_matrix.matrix
            array([[1, 0, 1],
                   [0, 1, 0],
                   [1, 0, 0]], dtype=int32)

            ```
        """
        _assert_supported_target(metric_target)
        _validate_input_tensors(predictions, targets, metric_target=metric_target)

        num_classes = len(classes)
        matrix: npt.NDArray[np.int32] = np.zeros(
            (num_classes + 1, num_classes + 1), dtype=np.int32
        )
        for true_batch, detection_batch in zip(targets, predictions):
            matrix += cls.evaluate_detection_batch(
                predictions=detection_batch,
                targets=true_batch,
                num_classes=num_classes,
                conf_threshold=conf_threshold,
                iou_threshold=iou_threshold,
                metric_target=metric_target,
            )
        return cls(
            matrix=matrix,
            classes=classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
            metric_target=metric_target,
        )

    @staticmethod
    def evaluate_detection_batch(
        predictions: npt.NDArray[np.float32],
        targets: npt.NDArray[np.float32],
        num_classes: int,
        conf_threshold: float,
        iou_threshold: float,
        metric_target: MetricTarget = MetricTarget.BOXES,
    ) -> npt.NDArray[np.int32]:
        """
        Calculate confusion matrix for a batch of detections for a single image.

        Args:
            predictions: Batch prediction. Describes a single image and
                has `shape = (M, 6)` or `shape = (M, 10)` depending on
                `metric_target`.
                If `MetricTarget.BOXES`, each row is in
                `(x_min, y_min, x_max, y_max, class, conf)` format.
                If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
                `(x1, y1, x2, y2, x3, y3, x4, y4, class, conf)` format.
            targets: Batch target labels. Describes a single image and
                has `shape = (N, 5)` or `shape = (N, 9)` depending on
                `metric_target`.
                If `MetricTarget.BOXES`, each row is in
                `(x_min, y_min, x_max, y_max, class)` format.
                If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
                `(x1, y1, x2, y2, x3, y3, x4, y4, class)` format.
            num_classes: Number of classes.
            conf_threshold: Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold: Detection iou threshold between `0` and `1`.
                Detections with lower iou will be classified as `FP`.
            metric_target: The type of detection data to use.
                Determines IoU function (`box_iou_batch` vs
                `oriented_box_iou_batch`) and coordinate column count.
                `MetricTarget.MASKS` is not supported.

        Returns:
            Confusion matrix based on a single image.
        """
        _assert_supported_target(metric_target)

        expected_pred_cols = (
            10 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 6
        )
        expected_target_cols = (
            9 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 5
        )
        if predictions.ndim != 2 or predictions.shape[1] != expected_pred_cols:
            raise ValueError(
                f"Predictions must have shape (M, {expected_pred_cols}). "
                f"Got {predictions.shape} instead."
            )
        if targets.ndim != 2 or targets.shape[1] != expected_target_cols:
            raise ValueError(
                f"Targets must have shape (N, {expected_target_cols}). "
                f"Got {targets.shape} instead."
            )

        result_matrix: npt.NDArray[np.int32] = np.zeros(
            (num_classes + 1, num_classes + 1), dtype=np.int32
        )

        # Filter predictions by confidence threshold
        coords_dim = 8 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 4
        class_id_idx = coords_dim
        conf_idx = coords_dim + 1

        confidence = predictions[:, conf_idx]
        detection_batch_filtered = predictions[confidence >= conf_threshold]

        if len(detection_batch_filtered) == 0:
            true_classes = _validated_class_ids(
                targets[:, class_id_idx], num_classes, "Target"
            )
            for gt_class in true_classes:
                result_matrix[gt_class, num_classes] += 1
            return result_matrix

        if len(targets) == 0:
            detection_classes = _validated_class_ids(
                detection_batch_filtered[:, class_id_idx], num_classes, "Prediction"
            )
            for det_class in detection_classes:
                result_matrix[num_classes, det_class] += 1
            return result_matrix

        true_classes = _validated_class_ids(
            targets[:, class_id_idx], num_classes, "Target"
        )
        detection_classes = _validated_class_ids(
            detection_batch_filtered[:, class_id_idx], num_classes, "Prediction"
        )
        true_boxes = targets[:, :coords_dim]
        detection_boxes = detection_batch_filtered[:, :coords_dim]

        # Calculate IoU matrix
        if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES:
            iou_batch = oriented_box_iou_batch(
                boxes_true=true_boxes, boxes_detection=detection_boxes
            )
        else:
            iou_batch = box_iou_batch(
                boxes_true=true_boxes, boxes_detection=detection_boxes
            )

        # Find all valid matches (IoU > threshold, regardless of class)
        # Use vectorized operations to avoid nested Python loops
        iou_mask = iou_batch > iou_threshold
        gt_indices, det_indices = np.nonzero(iou_mask)

        # If no pairs exceed the IoU threshold, skip matching
        if gt_indices.size == 0:
            valid_matches = []
        else:
            ious = iou_batch[gt_indices, det_indices]
            gt_match_classes = true_classes[gt_indices]
            det_match_classes = detection_classes[det_indices]
            class_matches = gt_match_classes == det_match_classes

            # Sort matches by class match first (True before False),
            # then by IoU descending.
            # np.lexsort sorts by the last key first, in ascending order.
            # We use ~class_matches so that True becomes 0
            # and False becomes 1 (True first),
            # and -ious so that larger IoUs come first.
            sort_indices = np.lexsort((-ious, ~class_matches))

            # Build list of matches in the same format as before:
            # (gt_idx, det_idx, iou, class_match)
            valid_matches = [
                (
                    int(gt_indices[idx]),
                    int(det_indices[idx]),
                    float(ious[idx]),
                    bool(class_matches[idx]),
                )
                for idx in sort_indices
            ]
        # Greedily assign matches, ensuring each GT
        # and detection is matched at most once
        matched_gt_idx = set()
        matched_det_idx = set()

        for gt_idx, det_idx, iou, class_match in valid_matches:
            if gt_idx not in matched_gt_idx and det_idx not in matched_det_idx:
                # Valid spatial match - record the class prediction
                gt_class = true_classes[gt_idx]
                det_class = detection_classes[det_idx]

                # This handles both correct classification (TP) and misclassification
                result_matrix[gt_class, det_class] += 1
                matched_gt_idx.add(gt_idx)
                matched_det_idx.add(det_idx)

        # Count unmatched ground truth as FN
        for gt_idx, gt_class in enumerate(true_classes):
            if gt_idx not in matched_gt_idx:
                result_matrix[gt_class, num_classes] += 1

        # Count unmatched detections as FP
        for det_idx, det_class in enumerate(detection_classes):
            if det_idx not in matched_det_idx:
                result_matrix[num_classes, det_class] += 1

        return result_matrix

    @staticmethod
    def _drop_extra_matches(
        matches: npt.NDArray[np.float32],
    ) -> npt.NDArray[np.float32]:
        """
        Deduplicate matches. If there are multiple matches for the same true or
        predicted box, only the one with the highest IoU is kept.
        """
        if matches.shape[0] > 0:
            matches = matches[matches[:, 2].argsort()[::-1]]
            matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
            matches = matches[matches[:, 2].argsort()[::-1]]
            matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
        result: npt.NDArray[np.float32] = matches
        return result

    @classmethod
    def benchmark(
        cls,
        dataset: DetectionDataset,
        callback: Callable[[npt.NDArray[np.uint8]], Detections],
        conf_threshold: float = 0.3,
        iou_threshold: float = 0.5,
        metric_target: MetricTarget = MetricTarget.BOXES,
        *,
        save_directory_path: str | Path | None = None,
    ) -> ConfusionMatrix:
        """
        Calculate confusion matrix from dataset and callback function.

        Args:
            dataset: Object detection dataset used for evaluation.
            callback: Function that takes an image as input and returns a
                Detections object.
            conf_threshold: Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold: Detection IoU threshold between `0` and `1`.
                Detections with lower IoU will be classified as `FP`.
            save_directory_path: Optional directory where per-image validation
                result grids are saved using the original image filenames. Images
                are written directly to this directory (no subdirectory is added).
                When ``None`` (default), no images are saved.
            metric_target: The type of detection data to use.
                Supports `MetricTarget.BOXES` and
                `MetricTarget.ORIENTED_BOUNDING_BOXES`. Passed through to
                `from_detections`. `MetricTarget.MASKS` is not supported.

        Returns:
            New instance of ConfusionMatrix.

        Example:
            ```python
            import supervision as sv
            from ultralytics import YOLO

            dataset = sv.DetectionDataset.from_yolo(...)

            model = YOLO(...)
            def callback(image: np.ndarray) -> sv.Detections:
                result = model(image)[0]
                return sv.Detections.from_ultralytics(result)

            confusion_matrix = sv.ConfusionMatrix.benchmark(
                dataset = dataset,
                callback = callback
            )

            print(confusion_matrix.matrix)
            # np.array([
            #     [0., 0., 0., 0.],
            #     [0., 1., 0., 1.],
            #     [0., 1., 1., 0.],
            #     [1., 1., 0., 0.]
            # ])
            ```
        """
        if save_directory_path is not None:
            save_directory = Path(save_directory_path)
            save_directory.mkdir(parents=True, exist_ok=True)

        predictions, targets = [], []
        for index, (image_name, image, annotation) in enumerate(dataset):
            predictions_batch = callback(image)
            predictions.append(predictions_batch)
            targets.append(annotation)

            if save_directory_path is not None:
                if isinstance(image_name, Path):
                    image_filename = image_name.name
                elif isinstance(image_name, str):
                    image_filename = Path(image_name).name
                else:
                    image_filename = f"image_{index:06d}.jpg"

                if Path(image_filename).suffix == "":
                    image_filename = f"{image_filename}.jpg"

                save_path = save_directory / image_filename
                if save_path.exists():
                    warnings.warn(
                        f"Validation image '{image_filename}' already exists at "
                        f"'{save_path}' and will be overwritten.",
                        UserWarning,
                        stacklevel=2,
                    )
                _save_detection_validation_visualization(
                    scene=image,
                    predictions=predictions_batch,
                    targets=annotation,
                    save_path=save_path,
                    conf_threshold=conf_threshold,
                    iou_threshold=iou_threshold,
                    class_names=dataset.classes,
                    metric_target=metric_target,
                )
        return cls.from_detections(
            predictions=predictions,
            targets=targets,
            classes=dataset.classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
            metric_target=metric_target,
        )

    def plot(
        self,
        save_path: str | None = None,
        title: str | None = None,
        classes: list[str] | None = None,
        normalize: bool = False,
        fig_size: tuple[int, int] = (12, 10),
    ) -> Figure:
        """
        Create confusion matrix plot and save it at selected location.

        Args:
            save_path: Path to save the plot. If not provided,
                plot will be displayed.
            title: Title of the plot.
            classes: List of classes to be displayed on the plot.
                If not provided, all classes will be displayed.
            normalize: If True, normalize the confusion matrix.
            fig_size: Size of the plot.

        Returns:
            Confusion matrix plot.
        """
        from matplotlib import pyplot as plt

        # Cast to float so that the NaN masking below never hits an integer
        # matrix (assigning NaN into an int array raises ValueError).
        array = self.matrix.astype(np.float64)

        if normalize:
            eps = 1e-8
            array = array / (array.sum(0).reshape(1, -1) + eps)

        array[array < 0.005] = np.nan

        fig, ax = plt.subplots(figsize=fig_size, tight_layout=True, facecolor="white")

        class_names = classes if classes is not None else self.classes
        use_labels_for_ticks = class_names is not None and (0 < len(class_names) < 99)
        if use_labels_for_ticks:
            x_tick_labels = [*class_names, "FN"]
            y_tick_labels = [*class_names, "FP"]
            num_ticks = len(x_tick_labels)
        else:
            x_tick_labels = None
            y_tick_labels = None
            num_ticks = len(array)
        im = ax.imshow(array, cmap="Blues")

        cbar = ax.figure.colorbar(im, ax=ax)
        cbar.mappable.set_clim(vmin=0, vmax=float(np.nanmax(array)))

        if x_tick_labels is None:
            tick_interval = 2
        else:
            tick_interval = 1
        ax.set_xticks(np.arange(0, num_ticks, tick_interval), labels=x_tick_labels)
        ax.set_yticks(np.arange(0, num_ticks, tick_interval), labels=y_tick_labels)

        plt.setp(ax.get_xticklabels(), rotation=90, ha="right", rotation_mode="default")

        labelsize = 10 if num_ticks < 50 else 8
        ax.tick_params(axis="both", which="both", labelsize=labelsize)

        if num_ticks < 30:
            for i in range(array.shape[0]):
                for j in range(array.shape[1]):
                    n_preds = array[i, j]
                    if not np.isnan(n_preds):
                        ax.text(
                            j,
                            i,
                            f"{n_preds:.2f}" if normalize else f"{n_preds:.0f}",
                            ha="center",
                            va="center",
                            color="black"
                            if n_preds < 0.5 * np.nanmax(array)
                            else "white",
                        )

        if title:
            ax.set_title(title, fontsize=20)

        ax.set_xlabel("Predicted")
        ax.set_ylabel("True")
        ax.set_facecolor("white")
        if save_path:
            fig.savefig(
                save_path, dpi=250, facecolor=fig.get_facecolor(), transparent=True
            )
        return fig

Methods:

benchmark(dataset: DetectionDataset, callback: Callable[[npt.NDArray[np.uint8]], Detections], conf_threshold: float = 0.3, iou_threshold: float = 0.5, metric_target: MetricTarget = MetricTarget.BOXES, *, save_directory_path: str | Path | None = None) -> ConfusionMatrix classmethod

Calculate confusion matrix from dataset and callback function.

Parameters:

Name	Type	Description	Default
dataset	DetectionDataset	Object detection dataset used for evaluation.	required
callback	Callable[[NDArray[uint8]], Detections]	Function that takes an image as input and returns a Detections object.	required
conf_threshold	float	Detection confidence threshold between 0 and 1. Detections with lower confidence will be excluded.	0.3
iou_threshold	float	Detection IoU threshold between 0 and 1. Detections with lower IoU will be classified as FP.	0.5
save_directory_path	str | Path | None	Optional directory where per-image validation result grids are saved using the original image filenames. Images are written directly to this directory (no subdirectory is added). When None (default), no images are saved.	None
metric_target	MetricTarget	The type of detection data to use. Supports MetricTarget.BOXES and MetricTarget.ORIENTED_BOUNDING_BOXES. Passed through to from_detections. MetricTarget.MASKS is not supported.	BOXES

Returns:

Type	Description
ConfusionMatrix	New instance of ConfusionMatrix.

Example

import supervision as sv
from ultralytics import YOLO

dataset = sv.DetectionDataset.from_yolo(...)

model = YOLO(...)
def callback(image: np.ndarray) -> sv.Detections:
    result = model(image)[0]
    return sv.Detections.from_ultralytics(result)

confusion_matrix = sv.ConfusionMatrix.benchmark(
    dataset = dataset,
    callback = callback
)

print(confusion_matrix.matrix)
# np.array([
#     [0., 0., 0., 0.],
#     [0., 1., 0., 1.],
#     [0., 1., 1., 0.],
#     [1., 1., 0., 0.]
# ])

Source code in src/supervision/metrics/detection.py

@classmethod
def benchmark(
    cls,
    dataset: DetectionDataset,
    callback: Callable[[npt.NDArray[np.uint8]], Detections],
    conf_threshold: float = 0.3,
    iou_threshold: float = 0.5,
    metric_target: MetricTarget = MetricTarget.BOXES,
    *,
    save_directory_path: str | Path | None = None,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix from dataset and callback function.

    Args:
        dataset: Object detection dataset used for evaluation.
        callback: Function that takes an image as input and returns a
            Detections object.
        conf_threshold: Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold: Detection IoU threshold between `0` and `1`.
            Detections with lower IoU will be classified as `FP`.
        save_directory_path: Optional directory where per-image validation
            result grids are saved using the original image filenames. Images
            are written directly to this directory (no subdirectory is added).
            When ``None`` (default), no images are saved.
        metric_target: The type of detection data to use.
            Supports `MetricTarget.BOXES` and
            `MetricTarget.ORIENTED_BOUNDING_BOXES`. Passed through to
            `from_detections`. `MetricTarget.MASKS` is not supported.

    Returns:
        New instance of ConfusionMatrix.

    Example:
        ```python
        import supervision as sv
        from ultralytics import YOLO

        dataset = sv.DetectionDataset.from_yolo(...)

        model = YOLO(...)
        def callback(image: np.ndarray) -> sv.Detections:
            result = model(image)[0]
            return sv.Detections.from_ultralytics(result)

        confusion_matrix = sv.ConfusionMatrix.benchmark(
            dataset = dataset,
            callback = callback
        )

        print(confusion_matrix.matrix)
        # np.array([
        #     [0., 0., 0., 0.],
        #     [0., 1., 0., 1.],
        #     [0., 1., 1., 0.],
        #     [1., 1., 0., 0.]
        # ])
        ```
    """
    if save_directory_path is not None:
        save_directory = Path(save_directory_path)
        save_directory.mkdir(parents=True, exist_ok=True)

    predictions, targets = [], []
    for index, (image_name, image, annotation) in enumerate(dataset):
        predictions_batch = callback(image)
        predictions.append(predictions_batch)
        targets.append(annotation)

        if save_directory_path is not None:
            if isinstance(image_name, Path):
                image_filename = image_name.name
            elif isinstance(image_name, str):
                image_filename = Path(image_name).name
            else:
                image_filename = f"image_{index:06d}.jpg"

            if Path(image_filename).suffix == "":
                image_filename = f"{image_filename}.jpg"

            save_path = save_directory / image_filename
            if save_path.exists():
                warnings.warn(
                    f"Validation image '{image_filename}' already exists at "
                    f"'{save_path}' and will be overwritten.",
                    UserWarning,
                    stacklevel=2,
                )
            _save_detection_validation_visualization(
                scene=image,
                predictions=predictions_batch,
                targets=annotation,
                save_path=save_path,
                conf_threshold=conf_threshold,
                iou_threshold=iou_threshold,
                class_names=dataset.classes,
                metric_target=metric_target,
            )
    return cls.from_detections(
        predictions=predictions,
        targets=targets,
        classes=dataset.classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
        metric_target=metric_target,
    )

evaluate_detection_batch(predictions: npt.NDArray[np.float32], targets: npt.NDArray[np.float32], num_classes: int, conf_threshold: float, iou_threshold: float, metric_target: MetricTarget = MetricTarget.BOXES) -> npt.NDArray[np.int32] staticmethod

Calculate confusion matrix for a batch of detections for a single image.

Parameters:

Name	Type	Description	Default
predictions	NDArray[float32]	Batch prediction. Describes a single image and has shape = (M, 6) or shape = (M, 10) depending on metric_target. If MetricTarget.BOXES, each row is in (x_min, y_min, x_max, y_max, class, conf) format. If MetricTarget.ORIENTED_BOUNDING_BOXES, each row is in (x1, y1, x2, y2, x3, y3, x4, y4, class, conf) format.	required
targets	NDArray[float32]	Batch target labels. Describes a single image and has shape = (N, 5) or shape = (N, 9) depending on metric_target. If MetricTarget.BOXES, each row is in (x_min, y_min, x_max, y_max, class) format. If MetricTarget.ORIENTED_BOUNDING_BOXES, each row is in (x1, y1, x2, y2, x3, y3, x4, y4, class) format.	required
num_classes	int	Number of classes.	required
conf_threshold	float	Detection confidence threshold between 0 and 1. Detections with lower confidence will be excluded.	required
iou_threshold	float	Detection iou threshold between 0 and 1. Detections with lower iou will be classified as FP.	required
metric_target	MetricTarget	The type of detection data to use. Determines IoU function (box_iou_batch vs oriented_box_iou_batch) and coordinate column count. MetricTarget.MASKS is not supported.	BOXES

Returns:

Type	Description
NDArray[int32]	Confusion matrix based on a single image.

Source code in src/supervision/metrics/detection.py

@staticmethod
def evaluate_detection_batch(
    predictions: npt.NDArray[np.float32],
    targets: npt.NDArray[np.float32],
    num_classes: int,
    conf_threshold: float,
    iou_threshold: float,
    metric_target: MetricTarget = MetricTarget.BOXES,
) -> npt.NDArray[np.int32]:
    """
    Calculate confusion matrix for a batch of detections for a single image.

    Args:
        predictions: Batch prediction. Describes a single image and
            has `shape = (M, 6)` or `shape = (M, 10)` depending on
            `metric_target`.
            If `MetricTarget.BOXES`, each row is in
            `(x_min, y_min, x_max, y_max, class, conf)` format.
            If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
            `(x1, y1, x2, y2, x3, y3, x4, y4, class, conf)` format.
        targets: Batch target labels. Describes a single image and
            has `shape = (N, 5)` or `shape = (N, 9)` depending on
            `metric_target`.
            If `MetricTarget.BOXES`, each row is in
            `(x_min, y_min, x_max, y_max, class)` format.
            If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
            `(x1, y1, x2, y2, x3, y3, x4, y4, class)` format.
        num_classes: Number of classes.
        conf_threshold: Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold: Detection iou threshold between `0` and `1`.
            Detections with lower iou will be classified as `FP`.
        metric_target: The type of detection data to use.
            Determines IoU function (`box_iou_batch` vs
            `oriented_box_iou_batch`) and coordinate column count.
            `MetricTarget.MASKS` is not supported.

    Returns:
        Confusion matrix based on a single image.
    """
    _assert_supported_target(metric_target)

    expected_pred_cols = (
        10 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 6
    )
    expected_target_cols = (
        9 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 5
    )
    if predictions.ndim != 2 or predictions.shape[1] != expected_pred_cols:
        raise ValueError(
            f"Predictions must have shape (M, {expected_pred_cols}). "
            f"Got {predictions.shape} instead."
        )
    if targets.ndim != 2 or targets.shape[1] != expected_target_cols:
        raise ValueError(
            f"Targets must have shape (N, {expected_target_cols}). "
            f"Got {targets.shape} instead."
        )

    result_matrix: npt.NDArray[np.int32] = np.zeros(
        (num_classes + 1, num_classes + 1), dtype=np.int32
    )

    # Filter predictions by confidence threshold
    coords_dim = 8 if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES else 4
    class_id_idx = coords_dim
    conf_idx = coords_dim + 1

    confidence = predictions[:, conf_idx]
    detection_batch_filtered = predictions[confidence >= conf_threshold]

    if len(detection_batch_filtered) == 0:
        true_classes = _validated_class_ids(
            targets[:, class_id_idx], num_classes, "Target"
        )
        for gt_class in true_classes:
            result_matrix[gt_class, num_classes] += 1
        return result_matrix

    if len(targets) == 0:
        detection_classes = _validated_class_ids(
            detection_batch_filtered[:, class_id_idx], num_classes, "Prediction"
        )
        for det_class in detection_classes:
            result_matrix[num_classes, det_class] += 1
        return result_matrix

    true_classes = _validated_class_ids(
        targets[:, class_id_idx], num_classes, "Target"
    )
    detection_classes = _validated_class_ids(
        detection_batch_filtered[:, class_id_idx], num_classes, "Prediction"
    )
    true_boxes = targets[:, :coords_dim]
    detection_boxes = detection_batch_filtered[:, :coords_dim]

    # Calculate IoU matrix
    if metric_target == MetricTarget.ORIENTED_BOUNDING_BOXES:
        iou_batch = oriented_box_iou_batch(
            boxes_true=true_boxes, boxes_detection=detection_boxes
        )
    else:
        iou_batch = box_iou_batch(
            boxes_true=true_boxes, boxes_detection=detection_boxes
        )

    # Find all valid matches (IoU > threshold, regardless of class)
    # Use vectorized operations to avoid nested Python loops
    iou_mask = iou_batch > iou_threshold
    gt_indices, det_indices = np.nonzero(iou_mask)

    # If no pairs exceed the IoU threshold, skip matching
    if gt_indices.size == 0:
        valid_matches = []
    else:
        ious = iou_batch[gt_indices, det_indices]
        gt_match_classes = true_classes[gt_indices]
        det_match_classes = detection_classes[det_indices]
        class_matches = gt_match_classes == det_match_classes

        # Sort matches by class match first (True before False),
        # then by IoU descending.
        # np.lexsort sorts by the last key first, in ascending order.
        # We use ~class_matches so that True becomes 0
        # and False becomes 1 (True first),
        # and -ious so that larger IoUs come first.
        sort_indices = np.lexsort((-ious, ~class_matches))

        # Build list of matches in the same format as before:
        # (gt_idx, det_idx, iou, class_match)
        valid_matches = [
            (
                int(gt_indices[idx]),
                int(det_indices[idx]),
                float(ious[idx]),
                bool(class_matches[idx]),
            )
            for idx in sort_indices
        ]
    # Greedily assign matches, ensuring each GT
    # and detection is matched at most once
    matched_gt_idx = set()
    matched_det_idx = set()

    for gt_idx, det_idx, iou, class_match in valid_matches:
        if gt_idx not in matched_gt_idx and det_idx not in matched_det_idx:
            # Valid spatial match - record the class prediction
            gt_class = true_classes[gt_idx]
            det_class = detection_classes[det_idx]

            # This handles both correct classification (TP) and misclassification
            result_matrix[gt_class, det_class] += 1
            matched_gt_idx.add(gt_idx)
            matched_det_idx.add(det_idx)

    # Count unmatched ground truth as FN
    for gt_idx, gt_class in enumerate(true_classes):
        if gt_idx not in matched_gt_idx:
            result_matrix[gt_class, num_classes] += 1

    # Count unmatched detections as FP
    for det_idx, det_class in enumerate(detection_classes):
        if det_idx not in matched_det_idx:
            result_matrix[num_classes, det_class] += 1

    return result_matrix

from_detections(predictions: list[Detections], targets: list[Detections], classes: list[str], conf_threshold: float = 0.3, iou_threshold: float = 0.5, metric_target: MetricTarget = MetricTarget.BOXES) -> ConfusionMatrix classmethod

Calculate confusion matrix based on predicted and ground-truth detections.

Parameters:

Name	Type	Description	Default
targets	list[Detections]	Detections objects from ground-truth.	required
predictions	list[Detections]	Detections objects predicted by the model.	required
classes	list[str]	Model class names.	required
conf_threshold	float	Detection confidence threshold between 0 and 1. Detections with lower confidence will be excluded.	0.3
iou_threshold	float	Detection IoU threshold between 0 and 1. Detections with lower IoU will be classified as FP.	0.5
metric_target	MetricTarget	The type of detection data to use. Supports MetricTarget.BOXES (default) and MetricTarget.ORIENTED_BOUNDING_BOXES. When using MetricTarget.ORIENTED_BOUNDING_BOXES, each Detections object must include OBB coordinates in detections.data[ORIENTED_BOX_COORDINATES] as a float32 array of shape (N, 8) (flat) or (N, 4, 2) (as stored by from_ultralytics); both are normalised to (N, 8) internally. MetricTarget.MASKS is not supported.	BOXES

Returns:

Type	Description
ConfusionMatrix	New instance of ConfusionMatrix.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> targets = [
...     sv.Detections(
...         xyxy=np.array([[0, 0, 10, 10], [50, 50, 60, 60]]),
...         class_id=np.array([0, 0])
...     )
... ]
>>> predictions = [
...     sv.Detections(
...         xyxy=np.array([[0, 0, 10, 10], [100, 100, 110, 110]]),
...         class_id=np.array([0, 0]),
...         confidence=np.array([0.9, 0.8])
...     )
... ]
>>> confusion_matrix = sv.ConfusionMatrix.from_detections(
...     predictions=predictions,
...     targets=targets,
...     classes=['person']
... )
>>> confusion_matrix.matrix
array([[1, 1],
       [1, 0]], dtype=int32)

Source code in src/supervision/metrics/detection.py

@classmethod
def from_detections(
    cls,
    predictions: list[Detections],
    targets: list[Detections],
    classes: list[str],
    conf_threshold: float = 0.3,
    iou_threshold: float = 0.5,
    metric_target: MetricTarget = MetricTarget.BOXES,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix based on predicted and ground-truth detections.

    Args:
        targets: Detections objects from ground-truth.
        predictions: Detections objects predicted by the model.
        classes: Model class names.
        conf_threshold: Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold: Detection IoU threshold between `0` and `1`.
            Detections with lower IoU will be classified as `FP`.
        metric_target: The type of detection data to use.
            Supports `MetricTarget.BOXES` (default) and
            `MetricTarget.ORIENTED_BOUNDING_BOXES`. When using
            `MetricTarget.ORIENTED_BOUNDING_BOXES`, each `Detections`
            object must include OBB coordinates in
            `detections.data[ORIENTED_BOX_COORDINATES]` as a float32
            array of shape `(N, 8)` (flat) or `(N, 4, 2)` (as stored by
            `from_ultralytics`); both are normalised to `(N, 8)` internally.
            `MetricTarget.MASKS` is not supported.

    Returns:
        New instance of ConfusionMatrix.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> targets = [
        ...     sv.Detections(
        ...         xyxy=np.array([[0, 0, 10, 10], [50, 50, 60, 60]]),
        ...         class_id=np.array([0, 0])
        ...     )
        ... ]
        >>> predictions = [
        ...     sv.Detections(
        ...         xyxy=np.array([[0, 0, 10, 10], [100, 100, 110, 110]]),
        ...         class_id=np.array([0, 0]),
        ...         confidence=np.array([0.9, 0.8])
        ...     )
        ... ]
        >>> confusion_matrix = sv.ConfusionMatrix.from_detections(
        ...     predictions=predictions,
        ...     targets=targets,
        ...     classes=['person']
        ... )
        >>> confusion_matrix.matrix
        array([[1, 1],
               [1, 0]], dtype=int32)

        ```
    """
    prediction_tensors = []
    target_tensors = []
    for prediction, target in zip(predictions, targets):
        prediction_tensors.append(
            detections_to_tensor(
                prediction, with_confidence=True, metric_target=metric_target
            )
        )
        target_tensors.append(
            detections_to_tensor(
                target, with_confidence=False, metric_target=metric_target
            )
        )
    return cls.from_tensors(
        predictions=prediction_tensors,
        targets=target_tensors,
        classes=classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
        metric_target=metric_target,
    )

from_tensors(predictions: list[npt.NDArray[np.float32]], targets: list[npt.NDArray[np.float32]], classes: list[str], conf_threshold: float = 0.3, iou_threshold: float = 0.5, metric_target: MetricTarget = MetricTarget.BOXES) -> ConfusionMatrix classmethod

Calculate confusion matrix based on predicted and ground-truth detections.

Parameters:

Name	Type	Description	Default
predictions	list[NDArray[float32]]	Each element of the list describes a single image and has shape = (M, 6) or shape = (M, 10) depending on metric_target. If MetricTarget.BOXES, each row is in (x_min, y_min, x_max, y_max, class, conf) format. If MetricTarget.ORIENTED_BOUNDING_BOXES, each row is in (x1, y1, x2, y2, x3, y3, x4, y4, class, conf) format.	required
targets	list[NDArray[float32]]	Each element of the list describes a single image and has shape = (N, 5) or shape = (N, 9) depending on metric_target. If MetricTarget.BOXES, each row is in (x_min, y_min, x_max, y_max, class) format. If MetricTarget.ORIENTED_BOUNDING_BOXES, each row is in (x1, y1, x2, y2, x3, y3, x4, y4, class) format.	required
classes	list[str]	Model class names.	required
conf_threshold	float	Detection confidence threshold between 0 and 1. Detections with lower confidence will be excluded.	0.3
iou_threshold	float	Detection iou threshold between 0 and 1. Detections with lower iou will be classified as FP.	0.5
metric_target	MetricTarget	The type of detection data to use. Determines expected tensor shapes (see Args above for column layouts). MetricTarget.MASKS is not supported.	BOXES

Returns:

Type	Description
ConfusionMatrix	New instance of ConfusionMatrix.

Examples:

>>> import supervision as sv
>>> import numpy as np
>>> targets = [
...     np.array([
...         [0.0, 0.0, 3.0, 3.0, 0],
...         [2.0, 2.0, 5.0, 5.0, 0],
...         [6.0, 1.0, 8.0, 3.0, 1],
...     ])
... ]
>>> predictions = [
...     np.array([
...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
...     ])
... ]
>>> confusion_matrix = sv.ConfusionMatrix.from_tensors(
...     predictions=predictions,
...     targets=targets,
...     classes=['person', 'dog']
... )
>>> confusion_matrix.matrix
array([[1, 0, 1],
       [0, 1, 0],
       [1, 0, 0]], dtype=int32)

Source code in src/supervision/metrics/detection.py

@classmethod
def from_tensors(
    cls,
    predictions: list[npt.NDArray[np.float32]],
    targets: list[npt.NDArray[np.float32]],
    classes: list[str],
    conf_threshold: float = 0.3,
    iou_threshold: float = 0.5,
    metric_target: MetricTarget = MetricTarget.BOXES,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix based on predicted and ground-truth detections.

    Args:
        predictions: Each element of the list describes a single
            image and has `shape = (M, 6)` or `shape = (M, 10)` depending on
            `metric_target`.
            If `MetricTarget.BOXES`, each row is in
            `(x_min, y_min, x_max, y_max, class, conf)` format.
            If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
            `(x1, y1, x2, y2, x3, y3, x4, y4, class, conf)` format.
        targets: Each element of the list describes a single
            image and has `shape = (N, 5)` or `shape = (N, 9)` depending on
            `metric_target`.
            If `MetricTarget.BOXES`, each row is in
            `(x_min, y_min, x_max, y_max, class)` format.
            If `MetricTarget.ORIENTED_BOUNDING_BOXES`, each row is in
            `(x1, y1, x2, y2, x3, y3, x4, y4, class)` format.
        classes: Model class names.
        conf_threshold: Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold: Detection iou threshold between `0` and `1`.
            Detections with lower iou will be classified as `FP`.
        metric_target: The type of detection data to use.
            Determines expected tensor shapes (see Args above for column
            layouts). `MetricTarget.MASKS` is not supported.

    Returns:
        New instance of ConfusionMatrix.

    Examples:
        ```pycon
        >>> import supervision as sv
        >>> import numpy as np
        >>> targets = [
        ...     np.array([
        ...         [0.0, 0.0, 3.0, 3.0, 0],
        ...         [2.0, 2.0, 5.0, 5.0, 0],
        ...         [6.0, 1.0, 8.0, 3.0, 1],
        ...     ])
        ... ]
        >>> predictions = [
        ...     np.array([
        ...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
        ...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
        ...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
        ...     ])
        ... ]
        >>> confusion_matrix = sv.ConfusionMatrix.from_tensors(
        ...     predictions=predictions,
        ...     targets=targets,
        ...     classes=['person', 'dog']
        ... )
        >>> confusion_matrix.matrix
        array([[1, 0, 1],
               [0, 1, 0],
               [1, 0, 0]], dtype=int32)

        ```
    """
    _assert_supported_target(metric_target)
    _validate_input_tensors(predictions, targets, metric_target=metric_target)

    num_classes = len(classes)
    matrix: npt.NDArray[np.int32] = np.zeros(
        (num_classes + 1, num_classes + 1), dtype=np.int32
    )
    for true_batch, detection_batch in zip(targets, predictions):
        matrix += cls.evaluate_detection_batch(
            predictions=detection_batch,
            targets=true_batch,
            num_classes=num_classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
            metric_target=metric_target,
        )
    return cls(
        matrix=matrix,
        classes=classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
        metric_target=metric_target,
    )

plot(save_path: str | None = None, title: str | None = None, classes: list[str] | None = None, normalize: bool = False, fig_size: tuple[int, int] = (12, 10)) -> Figure

Create confusion matrix plot and save it at selected location.

Parameters:

Name	Type	Description	Default
save_path	str | None	Path to save the plot. If not provided, plot will be displayed.	None
title	str | None	Title of the plot.	None
classes	list[str] | None	List of classes to be displayed on the plot. If not provided, all classes will be displayed.	None
normalize	bool	If True, normalize the confusion matrix.	False
fig_size	tuple[int, int]	Size of the plot.	(12, 10)

Returns:

Type	Description
Figure	Confusion matrix plot.

Source code in src/supervision/metrics/detection.py

def plot(
    self,
    save_path: str | None = None,
    title: str | None = None,
    classes: list[str] | None = None,
    normalize: bool = False,
    fig_size: tuple[int, int] = (12, 10),
) -> Figure:
    """
    Create confusion matrix plot and save it at selected location.

    Args:
        save_path: Path to save the plot. If not provided,
            plot will be displayed.
        title: Title of the plot.
        classes: List of classes to be displayed on the plot.
            If not provided, all classes will be displayed.
        normalize: If True, normalize the confusion matrix.
        fig_size: Size of the plot.

    Returns:
        Confusion matrix plot.
    """
    from matplotlib import pyplot as plt

    # Cast to float so that the NaN masking below never hits an integer
    # matrix (assigning NaN into an int array raises ValueError).
    array = self.matrix.astype(np.float64)

    if normalize:
        eps = 1e-8
        array = array / (array.sum(0).reshape(1, -1) + eps)

    array[array < 0.005] = np.nan

    fig, ax = plt.subplots(figsize=fig_size, tight_layout=True, facecolor="white")

    class_names = classes if classes is not None else self.classes
    use_labels_for_ticks = class_names is not None and (0 < len(class_names) < 99)
    if use_labels_for_ticks:
        x_tick_labels = [*class_names, "FN"]
        y_tick_labels = [*class_names, "FP"]
        num_ticks = len(x_tick_labels)
    else:
        x_tick_labels = None
        y_tick_labels = None
        num_ticks = len(array)
    im = ax.imshow(array, cmap="Blues")

    cbar = ax.figure.colorbar(im, ax=ax)
    cbar.mappable.set_clim(vmin=0, vmax=float(np.nanmax(array)))

    if x_tick_labels is None:
        tick_interval = 2
    else:
        tick_interval = 1
    ax.set_xticks(np.arange(0, num_ticks, tick_interval), labels=x_tick_labels)
    ax.set_yticks(np.arange(0, num_ticks, tick_interval), labels=y_tick_labels)

    plt.setp(ax.get_xticklabels(), rotation=90, ha="right", rotation_mode="default")

    labelsize = 10 if num_ticks < 50 else 8
    ax.tick_params(axis="both", which="both", labelsize=labelsize)

    if num_ticks < 30:
        for i in range(array.shape[0]):
            for j in range(array.shape[1]):
                n_preds = array[i, j]
                if not np.isnan(n_preds):
                    ax.text(
                        j,
                        i,
                        f"{n_preds:.2f}" if normalize else f"{n_preds:.0f}",
                        ha="center",
                        va="center",
                        color="black"
                        if n_preds < 0.5 * np.nanmax(array)
                        else "white",
                    )

    if title:
        ax.set_title(title, fontsize=20)

    ax.set_xlabel("Predicted")
    ax.set_ylabel("True")
    ax.set_facecolor("white")
    if save_path:
        fig.savefig(
            save_path, dpi=250, facecolor=fig.get_facecolor(), transparent=True
        )
    return fig

MeanAveragePrecision

supervision.metrics.detection.MeanAveragePrecision dataclass

Deprecated

MeanAveragePrecision is deprecated and will be removed in supervision-0.31.0.

The deprecated implementation provides results that are inconsistent with pycocotools. Please use supervision.metrics.mean_average_precision.MeanAveragePrecision instead, which matches the results of pycocotools and is now the recommended approach.

Mean Average Precision for object detection tasks.

Attributes:

Name	Type	Description
map50_95	float	Mean Average Precision (mAP) calculated over IoU thresholds ranging from 0.50 to 0.95 with a step size of 0.05.
map50	float	Mean Average Precision (mAP) calculated specifically at an IoU threshold of 0.50.
map75	float	Mean Average Precision (mAP) calculated specifically at an IoU threshold of 0.75.
per_class_ap50_95	NDArray[float64]	Average Precision (AP) values calculated over IoU thresholds ranging from 0.50 to 0.95 with a step size of 0.05, provided for each individual class.

Source code in src/supervision/metrics/detection.py

@deprecated_class(
    target=TargetMode.NOTIFY,
    deprecated_in="0.27.0",
    remove_in="0.31.0",
)
@dataclass(frozen=True)
class MeanAveragePrecision:
    """
    !!! deprecated "Deprecated"
        `MeanAveragePrecision` is **deprecated** and will be removed in
        `supervision-0.31.0`.

        The deprecated implementation provides results that are inconsistent with
        `pycocotools`. Please use
        `supervision.metrics.mean_average_precision.MeanAveragePrecision` instead,
        which matches the results of `pycocotools` and is now the recommended approach.

    Mean Average Precision for object detection tasks.

    Attributes:
        map50_95: Mean Average Precision (mAP) calculated over IoU thresholds
            ranging from `0.50` to `0.95` with a step size of `0.05`.
        map50: Mean Average Precision (mAP) calculated specifically at
            an IoU threshold of `0.50`.
        map75: Mean Average Precision (mAP) calculated specifically at
            an IoU threshold of `0.75`.
        per_class_ap50_95: Average Precision (AP) values calculated over
            IoU thresholds ranging from `0.50` to `0.95` with a step size of `0.05`,
            provided for each individual class.
    """

    map50_95: float
    map50: float
    map75: float
    per_class_ap50_95: npt.NDArray[np.float64]

    @classmethod
    def from_detections(
        cls,
        predictions: list[Detections],
        targets: list[Detections],
    ) -> MeanAveragePrecision:
        """
        Calculate mean average precision based on predicted and ground-truth detections.

        Args:
            targets: Detections objects from ground-truth.
            predictions: Detections objects predicted by the model.
        Returns:
            New instance of ConfusionMatrix.

        Examples:
            ```pycon
            >>> import numpy as np
            >>> import supervision as sv
            >>> targets = [
            ...     sv.Detections(
            ...         xyxy=np.array([[0, 0, 10, 10]]),
            ...         class_id=np.array([0])
            ...     )
            ... ]
            >>> predictions = [
            ...     sv.Detections(
            ...         xyxy=np.array([[0, 0, 10, 10]]),
            ...         class_id=np.array([0]),
            ...         confidence=np.array([0.9])
            ...     )
            ... ]
            >>> mAP = sv.MeanAveragePrecision.from_detections(
            ...     predictions=predictions,
            ...     targets=targets,
            ... )
            >>> round(float(mAP.map50), 2)
            1.0

            ```
        """
        prediction_tensors = []
        target_tensors = []
        for prediction, target in zip(predictions, targets):
            prediction_tensors.append(
                detections_to_tensor(prediction, with_confidence=True)
            )
            target_tensors.append(detections_to_tensor(target, with_confidence=False))
        return cls.from_tensors(
            predictions=prediction_tensors,
            targets=target_tensors,
        )

    @classmethod
    def benchmark(
        cls,
        dataset: DetectionDataset,
        callback: Callable[[npt.NDArray[np.uint8]], Detections],
    ) -> MeanAveragePrecision:
        """
        Calculate mean average precision from dataset and callback function.

        Args:
            dataset: Object detection dataset used for evaluation.
            callback: Function that takes
                an image as input and returns Detections object.
        Returns:
            New instance of MeanAveragePrecision.

        Example:
            ```python
            import supervision as sv
            from ultralytics import YOLO

            dataset = sv.DetectionDataset.from_yolo(...)

            model = YOLO(...)
            def callback(image: np.ndarray) -> sv.Detections:
                result = model(image)[0]
                return sv.Detections.from_ultralytics(result)

            mean_average_precision = sv.MeanAveragePrecision.benchmark(
                dataset = dataset,
                callback = callback
            )

            print(mean_average_precision.map50_95)
            # 0.433
            ```
        """
        predictions, targets = [], []
        for _, image, annotation in dataset:
            predictions_batch = callback(image)
            predictions.append(predictions_batch)
            targets.append(annotation)
        return cls.from_detections(
            predictions=predictions,
            targets=targets,
        )

    @classmethod
    def from_tensors(
        cls,
        predictions: list[npt.NDArray[np.float32]],
        targets: list[npt.NDArray[np.float32]],
    ) -> MeanAveragePrecision:
        """
        Calculate Mean Average Precision based on predicted and ground-truth
            detections at different threshold.

        Args:
            predictions: Each element of the list describes
                a single image and has `shape = (M, 6)` where `M` is
                the number of detected objects. Each row is expected to be
                in `(x_min, y_min, x_max, y_max, class, conf)` format.
            targets: Each element of the list describes a single
                image and has `shape = (N, 5)` where `N` is the
                number of ground-truth objects. Each row is expected to be in
                `(x_min, y_min, x_max, y_max, class)` format. An empty array
                (``N = 0``) represents a background image; all predictions on
                that image count as false positives and reduce AP accordingly.

        Returns:
            MeanAveragePrecision: New instance computed from the provided
                predictions and targets.

        Examples:
            ```pycon
            >>> import supervision as sv
            >>> import numpy as np
            >>> targets = [
            ...     np.array([
            ...         [0.0, 0.0, 3.0, 3.0, 0],
            ...         [2.0, 2.0, 5.0, 5.0, 0],
            ...         [6.0, 1.0, 8.0, 3.0, 1],
            ...     ])
            ... ]
            >>> predictions = [
            ...     np.array([
            ...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
            ...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
            ...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
            ...     ])
            ... ]
            >>> mAP = sv.MeanAveragePrecision.from_tensors(
            ...     predictions=predictions,
            ...     targets=targets,
            ... )
            >>> round(float(mAP.map50), 2)
            0.75

            >>> bg_pred = [np.array([[0., 0., 10., 10., 0, 0.9]], dtype=np.float32)]
            >>> bg_tgt = [np.zeros((0, 5), dtype=np.float32)]
            >>> mAP_bg = sv.MeanAveragePrecision.from_tensors(
            ...     predictions=bg_pred,
            ...     targets=bg_tgt,
            ... )
            >>> float(mAP_bg.map50)
            0.0

            ```
        """
        _validate_input_tensors(predictions, targets)
        iou_thresholds = np.linspace(0.5, 0.95, 10, dtype=np.float32)
        stats = []

        # Gather matching stats for predictions and targets
        for true_objs, predicted_objs in zip(targets, predictions):
            if predicted_objs.shape[0] == 0:
                if true_objs.shape[0]:
                    stats.append(
                        (
                            np.zeros((0, iou_thresholds.size), dtype=bool),
                            *np.zeros((2, 0)),
                            true_objs[:, 4],
                        )
                    )
                continue

            if true_objs.shape[0]:
                matches = cls._match_detection_batch(
                    predicted_objs, true_objs, iou_thresholds
                )
                stats.append(
                    (
                        matches,
                        predicted_objs[:, 5],
                        predicted_objs[:, 4],
                        true_objs[:, 4],
                    )
                )
            else:
                # Background image: no GT boxes, so all predictions are FP matches.
                # This lowers AP for classes that appear in at least one GT image
                # elsewhere; classes absent from all GT images are excluded from AP
                # (they never appear in true_class_ids and are skipped by the AP loop).
                stats.append(
                    (
                        np.zeros(
                            (predicted_objs.shape[0], iou_thresholds.size), dtype=bool
                        ),
                        predicted_objs[:, 5],
                        predicted_objs[:, 4],
                        np.zeros((0,), dtype=np.float32),
                    )
                )

        # Compute average precisions if any matches exist
        if stats:
            concatenated_stats = [np.concatenate(items, 0) for items in zip(*stats)]
            average_precisions = cls._average_precisions_per_class(
                cast(npt.NDArray[np.bool_], concatenated_stats[0]),
                cast(npt.NDArray[np.float32], concatenated_stats[1]),
                cast(npt.NDArray[np.int32], concatenated_stats[2]),
                cast(npt.NDArray[np.int32], concatenated_stats[3]),
            )
            if average_precisions.size == 0:
                # All images had no ground-truth objects; FPs recorded but no class
                # to accumulate AP over → return 0.0 rather than NaN.
                map50, map75, map50_95 = 0.0, 0.0, 0.0
            else:
                map50 = float(average_precisions[:, 0].mean())
                map75 = float(average_precisions[:, 5].mean())
                map50_95 = float(average_precisions.mean())
        else:
            map50, map75, map50_95 = 0, 0, 0
            average_precisions = np.array([])

        return cls(
            map50_95=map50_95,
            map50=map50,
            map75=map75,
            per_class_ap50_95=average_precisions,
        )

    @staticmethod
    def compute_average_precision(
        recall: npt.NDArray[np.float64],
        precision: npt.NDArray[np.float64],
    ) -> float:
        """
        Compute the average precision using 101-point interpolation (COCO), given
            the recall and precision curves.

        Args:
            recall: The recall curve.
            precision: The precision curve.

        Returns:
            Average precision.
        """
        extended_recall = np.concatenate(([0.0], recall, [1.0]))
        extended_precision = np.concatenate(([0.0], precision, [0.0]))
        max_accumulated_precision = np.flip(
            np.maximum.accumulate(np.flip(extended_precision))
        )
        interpolated_recall_levels = np.linspace(0, 1, 101)
        recall_indices = np.searchsorted(
            extended_recall, interpolated_recall_levels, side="left"
        )
        interpolated_precision = max_accumulated_precision[recall_indices]
        average_precision = np.mean(interpolated_precision)

        return float(average_precision)

    @staticmethod
    def _match_detection_batch(
        predictions: npt.NDArray[np.float32],
        targets: npt.NDArray[np.float32],
        iou_thresholds: npt.NDArray[np.float32],
    ) -> npt.NDArray[np.bool_]:
        """
        Match predictions with target labels based on IoU levels.

        Args:
            predictions: Batch prediction. Describes a single image and
                has `shape = (M, 6)` where `M` is the number of detected objects.
                Each row is expected to be in
                `(x_min, y_min, x_max, y_max, class, conf)` format.
            targets: Batch target labels. Describes a single image and
                has `shape = (N, 5)` where `N` is the number of ground-truth objects.
                Each row is expected to be in
                `(x_min, y_min, x_max, y_max, class)` format.
            iou_thresholds: Array contains different IoU thresholds.

        Returns:
            Matched prediction with target labels result.
        """
        num_predictions, num_iou_levels = predictions.shape[0], iou_thresholds.shape[0]
        correct = np.zeros((num_predictions, num_iou_levels), dtype=bool)
        iou = box_iou_batch(targets[:, :4], predictions[:, :4])
        correct_class = targets[:, 4:5] == predictions[:, 4]

        for i, iou_level in enumerate(iou_thresholds):
            matched_indices = np.where((iou >= iou_level) & correct_class)

            for t, p in _greedy_match(iou, matched_indices):
                correct[p, i] = True
        result: npt.NDArray[np.bool_] = correct
        return result

    @staticmethod
    def _average_precisions_per_class(
        matches: npt.NDArray[np.bool_],
        prediction_confidence: npt.NDArray[np.float32],
        prediction_class_ids: npt.NDArray[np.int32],
        true_class_ids: npt.NDArray[np.int32],
        eps: float = 1e-16,
    ) -> npt.NDArray[np.float64]:
        """
        Compute the average precision, given the recall and precision curves.
        Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.

        Args:
            matches: True positives.
            prediction_confidence: Objectness value from 0-1.
            prediction_class_ids: Predicted object classes.
            true_class_ids: True object classes.
            eps: Small value to prevent division by zero.

        Returns:
            Average precision for different IoU levels.
        """
        sorted_indices = np.argsort(-prediction_confidence)
        matches = matches[sorted_indices]
        prediction_class_ids = prediction_class_ids[sorted_indices]

        unique_classes, class_counts = np.unique(true_class_ids, return_counts=True)
        num_classes = unique_classes.shape[0]

        average_precisions: npt.NDArray[np.float64] = np.zeros(
            (num_classes, matches.shape[1]), dtype=np.float64
        )

        for class_idx, class_id in enumerate(unique_classes):
            is_class = prediction_class_ids == class_id
            total_true = class_counts[class_idx]
            total_prediction = is_class.sum()

            if total_prediction == 0 or total_true == 0:
                continue

            false_positives = (1 - matches[is_class]).cumsum(0)
            true_positives = matches[is_class].cumsum(0)
            recall = true_positives / (total_true + eps)
            precision = true_positives / (true_positives + false_positives)

            for iou_level_idx in range(matches.shape[1]):
                average_precisions[class_idx, iou_level_idx] = (
                    MeanAveragePrecision.compute_average_precision(
                        recall[:, iou_level_idx], precision[:, iou_level_idx]
                    )
                )

        result: npt.NDArray[np.float64] = average_precisions
        return result

Methods:

benchmark(dataset: DetectionDataset, callback: Callable[[npt.NDArray[np.uint8]], Detections]) -> MeanAveragePrecision classmethod

Calculate mean average precision from dataset and callback function.

Parameters:

Name	Type	Description	Default
dataset	DetectionDataset	Object detection dataset used for evaluation.	required
callback	Callable[[NDArray[uint8]], Detections]	Function that takes an image as input and returns Detections object.	required

Returns: New instance of MeanAveragePrecision.

Example

import supervision as sv
from ultralytics import YOLO

dataset = sv.DetectionDataset.from_yolo(...)

model = YOLO(...)
def callback(image: np.ndarray) -> sv.Detections:
    result = model(image)[0]
    return sv.Detections.from_ultralytics(result)

mean_average_precision = sv.MeanAveragePrecision.benchmark(
    dataset = dataset,
    callback = callback
)

print(mean_average_precision.map50_95)
# 0.433

Source code in src/supervision/metrics/detection.py

@classmethod
def benchmark(
    cls,
    dataset: DetectionDataset,
    callback: Callable[[npt.NDArray[np.uint8]], Detections],
) -> MeanAveragePrecision:
    """
    Calculate mean average precision from dataset and callback function.

    Args:
        dataset: Object detection dataset used for evaluation.
        callback: Function that takes
            an image as input and returns Detections object.
    Returns:
        New instance of MeanAveragePrecision.

    Example:
        ```python
        import supervision as sv
        from ultralytics import YOLO

        dataset = sv.DetectionDataset.from_yolo(...)

        model = YOLO(...)
        def callback(image: np.ndarray) -> sv.Detections:
            result = model(image)[0]
            return sv.Detections.from_ultralytics(result)

        mean_average_precision = sv.MeanAveragePrecision.benchmark(
            dataset = dataset,
            callback = callback
        )

        print(mean_average_precision.map50_95)
        # 0.433
        ```
    """
    predictions, targets = [], []
    for _, image, annotation in dataset:
        predictions_batch = callback(image)
        predictions.append(predictions_batch)
        targets.append(annotation)
    return cls.from_detections(
        predictions=predictions,
        targets=targets,
    )

compute_average_precision(recall: npt.NDArray[np.float64], precision: npt.NDArray[np.float64]) -> float staticmethod

Compute the average precision using 101-point interpolation (COCO), given the recall and precision curves.

Parameters:

Name	Type	Description	Default
recall	NDArray[float64]	The recall curve.	required
precision	NDArray[float64]	The precision curve.	required

Returns:

Type	Description
float	Average precision.

Source code in src/supervision/metrics/detection.py

@staticmethod
def compute_average_precision(
    recall: npt.NDArray[np.float64],
    precision: npt.NDArray[np.float64],
) -> float:
    """
    Compute the average precision using 101-point interpolation (COCO), given
        the recall and precision curves.

    Args:
        recall: The recall curve.
        precision: The precision curve.

    Returns:
        Average precision.
    """
    extended_recall = np.concatenate(([0.0], recall, [1.0]))
    extended_precision = np.concatenate(([0.0], precision, [0.0]))
    max_accumulated_precision = np.flip(
        np.maximum.accumulate(np.flip(extended_precision))
    )
    interpolated_recall_levels = np.linspace(0, 1, 101)
    recall_indices = np.searchsorted(
        extended_recall, interpolated_recall_levels, side="left"
    )
    interpolated_precision = max_accumulated_precision[recall_indices]
    average_precision = np.mean(interpolated_precision)

    return float(average_precision)

from_detections(predictions: list[Detections], targets: list[Detections]) -> MeanAveragePrecision classmethod

Calculate mean average precision based on predicted and ground-truth detections.

Parameters:

Name	Type	Description	Default
targets	list[Detections]	Detections objects from ground-truth.	required
predictions	list[Detections]	Detections objects predicted by the model.	required

Returns: New instance of ConfusionMatrix.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> targets = [
...     sv.Detections(
...         xyxy=np.array([[0, 0, 10, 10]]),
...         class_id=np.array([0])
...     )
... ]
>>> predictions = [
...     sv.Detections(
...         xyxy=np.array([[0, 0, 10, 10]]),
...         class_id=np.array([0]),
...         confidence=np.array([0.9])
...     )
... ]
>>> mAP = sv.MeanAveragePrecision.from_detections(
...     predictions=predictions,
...     targets=targets,
... )
>>> round(float(mAP.map50), 2)
1.0

Source code in src/supervision/metrics/detection.py

@classmethod
def from_detections(
    cls,
    predictions: list[Detections],
    targets: list[Detections],
) -> MeanAveragePrecision:
    """
    Calculate mean average precision based on predicted and ground-truth detections.

    Args:
        targets: Detections objects from ground-truth.
        predictions: Detections objects predicted by the model.
    Returns:
        New instance of ConfusionMatrix.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> targets = [
        ...     sv.Detections(
        ...         xyxy=np.array([[0, 0, 10, 10]]),
        ...         class_id=np.array([0])
        ...     )
        ... ]
        >>> predictions = [
        ...     sv.Detections(
        ...         xyxy=np.array([[0, 0, 10, 10]]),
        ...         class_id=np.array([0]),
        ...         confidence=np.array([0.9])
        ...     )
        ... ]
        >>> mAP = sv.MeanAveragePrecision.from_detections(
        ...     predictions=predictions,
        ...     targets=targets,
        ... )
        >>> round(float(mAP.map50), 2)
        1.0

        ```
    """
    prediction_tensors = []
    target_tensors = []
    for prediction, target in zip(predictions, targets):
        prediction_tensors.append(
            detections_to_tensor(prediction, with_confidence=True)
        )
        target_tensors.append(detections_to_tensor(target, with_confidence=False))
    return cls.from_tensors(
        predictions=prediction_tensors,
        targets=target_tensors,
    )

from_tensors(predictions: list[npt.NDArray[np.float32]], targets: list[npt.NDArray[np.float32]]) -> MeanAveragePrecision classmethod

Calculate Mean Average Precision based on predicted and ground-truth detections at different threshold.

Parameters:

Name	Type	Description	Default
predictions	list[NDArray[float32]]	Each element of the list describes a single image and has shape = (M, 6) where M is the number of detected objects. Each row is expected to be in (x_min, y_min, x_max, y_max, class, conf) format.	required
targets	list[NDArray[float32]]	Each element of the list describes a single image and has shape = (N, 5) where N is the number of ground-truth objects. Each row is expected to be in (x_min, y_min, x_max, y_max, class) format. An empty array (N = 0) represents a background image; all predictions on that image count as false positives and reduce AP accordingly.	required

Returns:

Name	Type	Description
MeanAveragePrecision	MeanAveragePrecision	New instance computed from the provided predictions and targets.

Examples:

>>> import supervision as sv
>>> import numpy as np
>>> targets = [
...     np.array([
...         [0.0, 0.0, 3.0, 3.0, 0],
...         [2.0, 2.0, 5.0, 5.0, 0],
...         [6.0, 1.0, 8.0, 3.0, 1],
...     ])
... ]
>>> predictions = [
...     np.array([
...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
...     ])
... ]
>>> mAP = sv.MeanAveragePrecision.from_tensors(
...     predictions=predictions,
...     targets=targets,
... )
>>> round(float(mAP.map50), 2)
0.75

>>> bg_pred = [np.array([[0., 0., 10., 10., 0, 0.9]], dtype=np.float32)]
>>> bg_tgt = [np.zeros((0, 5), dtype=np.float32)]
>>> mAP_bg = sv.MeanAveragePrecision.from_tensors(
...     predictions=bg_pred,
...     targets=bg_tgt,
... )
>>> float(mAP_bg.map50)
0.0

Source code in src/supervision/metrics/detection.py

@classmethod
def from_tensors(
    cls,
    predictions: list[npt.NDArray[np.float32]],
    targets: list[npt.NDArray[np.float32]],
) -> MeanAveragePrecision:
    """
    Calculate Mean Average Precision based on predicted and ground-truth
        detections at different threshold.

    Args:
        predictions: Each element of the list describes
            a single image and has `shape = (M, 6)` where `M` is
            the number of detected objects. Each row is expected to be
            in `(x_min, y_min, x_max, y_max, class, conf)` format.
        targets: Each element of the list describes a single
            image and has `shape = (N, 5)` where `N` is the
            number of ground-truth objects. Each row is expected to be in
            `(x_min, y_min, x_max, y_max, class)` format. An empty array
            (``N = 0``) represents a background image; all predictions on
            that image count as false positives and reduce AP accordingly.

    Returns:
        MeanAveragePrecision: New instance computed from the provided
            predictions and targets.

    Examples:
        ```pycon
        >>> import supervision as sv
        >>> import numpy as np
        >>> targets = [
        ...     np.array([
        ...         [0.0, 0.0, 3.0, 3.0, 0],
        ...         [2.0, 2.0, 5.0, 5.0, 0],
        ...         [6.0, 1.0, 8.0, 3.0, 1],
        ...     ])
        ... ]
        >>> predictions = [
        ...     np.array([
        ...         [0.0, 0.0, 3.0, 3.0, 0, 0.9],
        ...         [0.1, 0.1, 3.0, 3.0, 0, 0.9],
        ...         [6.0, 1.0, 8.0, 3.0, 1, 0.8],
        ...     ])
        ... ]
        >>> mAP = sv.MeanAveragePrecision.from_tensors(
        ...     predictions=predictions,
        ...     targets=targets,
        ... )
        >>> round(float(mAP.map50), 2)
        0.75

        >>> bg_pred = [np.array([[0., 0., 10., 10., 0, 0.9]], dtype=np.float32)]
        >>> bg_tgt = [np.zeros((0, 5), dtype=np.float32)]
        >>> mAP_bg = sv.MeanAveragePrecision.from_tensors(
        ...     predictions=bg_pred,
        ...     targets=bg_tgt,
        ... )
        >>> float(mAP_bg.map50)
        0.0

        ```
    """
    _validate_input_tensors(predictions, targets)
    iou_thresholds = np.linspace(0.5, 0.95, 10, dtype=np.float32)
    stats = []

    # Gather matching stats for predictions and targets
    for true_objs, predicted_objs in zip(targets, predictions):
        if predicted_objs.shape[0] == 0:
            if true_objs.shape[0]:
                stats.append(
                    (
                        np.zeros((0, iou_thresholds.size), dtype=bool),
                        *np.zeros((2, 0)),
                        true_objs[:, 4],
                    )
                )
            continue

        if true_objs.shape[0]:
            matches = cls._match_detection_batch(
                predicted_objs, true_objs, iou_thresholds
            )
            stats.append(
                (
                    matches,
                    predicted_objs[:, 5],
                    predicted_objs[:, 4],
                    true_objs[:, 4],
                )
            )
        else:
            # Background image: no GT boxes, so all predictions are FP matches.
            # This lowers AP for classes that appear in at least one GT image
            # elsewhere; classes absent from all GT images are excluded from AP
            # (they never appear in true_class_ids and are skipped by the AP loop).
            stats.append(
                (
                    np.zeros(
                        (predicted_objs.shape[0], iou_thresholds.size), dtype=bool
                    ),
                    predicted_objs[:, 5],
                    predicted_objs[:, 4],
                    np.zeros((0,), dtype=np.float32),
                )
            )

    # Compute average precisions if any matches exist
    if stats:
        concatenated_stats = [np.concatenate(items, 0) for items in zip(*stats)]
        average_precisions = cls._average_precisions_per_class(
            cast(npt.NDArray[np.bool_], concatenated_stats[0]),
            cast(npt.NDArray[np.float32], concatenated_stats[1]),
            cast(npt.NDArray[np.int32], concatenated_stats[2]),
            cast(npt.NDArray[np.int32], concatenated_stats[3]),
        )
        if average_precisions.size == 0:
            # All images had no ground-truth objects; FPs recorded but no class
            # to accumulate AP over → return 0.0 rather than NaN.
            map50, map75, map50_95 = 0.0, 0.0, 0.0
        else:
            map50 = float(average_precisions[:, 0].mean())
            map75 = float(average_precisions[:, 5].mean())
            map50_95 = float(average_precisions.mean())
    else:
        map50, map75, map50_95 = 0, 0, 0
        average_precisions = np.array([])

    return cls(
        map50_95=map50_95,
        map50=map50,
        map75=map75,
        per_class_ap50_95=average_precisions,
    )

Comments