Reference for ultralytics/utils/metrics.py

class ConfusionMatrix(DataExportMixin):
    """A class for calculating and updating a confusion matrix for object detection and classification tasks.

    Attributes:
        task (str): The type of task, either 'detect' or 'classify'.
        matrix (np.ndarray): The confusion matrix, with dimensions depending on the task.
        nc (int): The number of classes.
        names (dict[int, str]): The names of the classes, used as labels on the plot.
        matches (dict | None): Contains the indices of ground truths and predictions categorized into TP, FP and FN.
    """

    def __init__(self, names: dict[int, str] = {}, task: str = "detect", save_matches: bool = False):
        """Initialize a ConfusionMatrix instance.

        Args:
            names (dict[int, str], optional): Names of classes, used as labels on the plot.
            task (str, optional): Type of task, either 'detect' or 'classify'.
            save_matches (bool, optional): Save the indices of GTs, TPs, FPs, FNs for visualization.
        """
        self.task = task
        self.nc = len(names)  # number of classes
        self.matrix = (
            np.zeros((self.nc, self.nc))
            if self.task in {"classify", "semantic"}
            else np.zeros((self.nc + 1, self.nc + 1))
        )
        self.names = names  # name of classes
        self.matches = {} if save_matches else None

def _append_matches(self, mtype: str, batch: dict[str, Any], idx: int) -> None:
    """Append the matches to TP, FP, FN or GT list for the last batch.

    This method updates the matches dictionary by appending specific batch data to the appropriate match type (True
    Positive, False Positive, or False Negative).

    Args:
        mtype (str): Match type identifier ('TP', 'FP', 'FN' or 'GT').
        batch (dict[str, Any]): Batch data containing detection results with keys like 'bboxes', 'cls', 'conf',
            'keypoints', 'masks'.
        idx (int): Index of the specific detection to append from the batch.

    Notes:
        For masks, handles both overlap and non-overlap cases. When masks.max() > 1.0, it indicates
        overlap_mask=True with shape (1, H, W), otherwise uses direct indexing.
    """
    if self.matches is None:
        return
    for k, v in batch.items():
        if k in {"bboxes", "cls", "conf", "keypoints"}:
            self.matches[mtype][k] += v[[idx]]
        elif k == "masks":
            # NOTE: masks.max() > 1.0 means overlap_mask=True with (1, H, W) shape
            self.matches[mtype][k] += [v[0] == idx + 1] if v.max() > 1.0 else [v[idx]]

def matrix(self):
    """Return the confusion matrix."""
    return self.matrix

@TryExcept(msg="ConfusionMatrix plot failure")
@plt_settings()
def plot(self, normalize: bool = True, save_dir: str = "", on_plot=None):
    """Plot the confusion matrix using matplotlib and save it to a file.

    Args:
        normalize (bool, optional): Whether to normalize the confusion matrix.
        save_dir (str, optional): Directory where the plot will be saved.
        on_plot (callable, optional): An optional callback to pass plots path and data when they are rendered.
    """
    import matplotlib.pyplot as plt  # scope for faster 'import ultralytics'

    array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1e-9) if normalize else 1)  # normalize columns
    array[array < 0.005] = np.nan  # don't annotate (would appear as 0.00)

    fig, ax = plt.subplots(1, 1, figsize=(12, 9))
    names, n = list(self.names.values()), self.nc
    if self.nc >= 100:  # downsample for large class count
        k = max(2, self.nc // 60)  # step size for downsampling, always > 1
        keep_idx = slice(None, None, k)  # create slice instead of array
        names = names[keep_idx]  # slice class names
        array = array[keep_idx, :][:, keep_idx]  # slice matrix rows and cols
        n = (self.nc + k - 1) // k  # number of retained classes
    nc = n if self.task == "classify" else n + 1  # adjust for background if needed
    ticklabels = "auto"
    if 0 < nc < 99:
        ticklabels = names if self.task in {"classify", "semantic"} else [*names, "background"]
    xy_ticks = np.arange(len(ticklabels)) if ticklabels != "auto" else np.arange(nc)
    tick_fontsize = max(6, 15 - 0.1 * nc)  # Minimum size is 6
    label_fontsize = max(6, 12 - 0.1 * nc)
    title_fontsize = max(6, 12 - 0.1 * nc)
    btm = max(0.1, 0.25 - 0.001 * nc)  # Minimum value is 0.1
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")  # suppress empty matrix RuntimeWarning: All-NaN slice encountered
        im = ax.imshow(array, cmap="Blues", vmin=0.0, interpolation="none")
        ax.xaxis.set_label_position("bottom")
        if nc < 30:  # Add score for each cell of confusion matrix
            color_threshold = 0.45 * (1 if normalize else np.nanmax(array))  # text color threshold
            for i, row in enumerate(array[:nc]):
                for j, val in enumerate(row[:nc]):
                    val = array[i, j]
                    if np.isnan(val):
                        continue
                    ax.text(
                        j,
                        i,
                        f"{val:.2f}" if normalize else f"{int(val)}",
                        ha="center",
                        va="center",
                        fontsize=10,
                        color="white" if val > color_threshold else "black",
                    )
        cbar = fig.colorbar(im, ax=ax, fraction=0.046, pad=0.05)
    title = "Confusion Matrix" + " Normalized" * normalize
    ax.set_xlabel("True", fontsize=label_fontsize, labelpad=10)
    ax.set_ylabel("Predicted", fontsize=label_fontsize, labelpad=10)
    ax.set_title(title, fontsize=title_fontsize, pad=20)
    ax.set_xticks(xy_ticks)
    ax.set_yticks(xy_ticks)
    ax.tick_params(axis="x", bottom=True, top=False, labelbottom=True, labeltop=False)
    ax.tick_params(axis="y", left=True, right=False, labelleft=True, labelright=False)
    if ticklabels != "auto":
        ax.set_xticklabels(ticklabels, fontsize=tick_fontsize, rotation=90, ha="center")
        ax.set_yticklabels(ticklabels, fontsize=tick_fontsize)
    for s in {"left", "right", "bottom", "top", "outline"}:
        if s != "outline":
            ax.spines[s].set_visible(False)  # Confusion matrix plot don't have outline
        cbar.ax.spines[s].set_visible(False)
    fig.subplots_adjust(left=0, right=0.84, top=0.94, bottom=btm)  # Adjust layout to ensure equal margins
    plot_fname = Path(save_dir) / f"{title.lower().replace(' ', '_')}.png"
    fig.savefig(plot_fname, dpi=250)
    plt.close(fig)
    if on_plot:
        on_plot(plot_fname, {"type": "confusion_matrix", "matrix": self.matrix.tolist()})

def plot_matches(
    self, img: torch.Tensor, im_file: str, save_dir: Path, show_labels: bool = True, show_conf: bool = True
) -> None:
    """Plot grid of GT, TP, FP, FN for each image.

    Args:
        img (torch.Tensor): Image to plot onto.
        im_file (str): Image filename to save visualizations.
        save_dir (Path): Location to save the visualizations to.
        show_labels (bool): Whether to display class labels in the visualization.
        show_conf (bool): Whether to display confidence values in the visualization.
    """
    if not self.matches:
        return
    from .ops import xyxy2xywh
    from .plotting import plot_images

    # Create batch of 4 (GT, TP, FP, FN)
    labels = defaultdict(list)
    for i, mtype in enumerate(["GT", "FP", "TP", "FN"]):
        mbatch = self.matches[mtype]
        if "conf" not in mbatch:
            mbatch["conf"] = torch.tensor([1.0] * len(mbatch["bboxes"]), device=img.device)
        mbatch["batch_idx"] = torch.ones(len(mbatch["bboxes"]), device=img.device) * i
        for k in mbatch:
            labels[k] += mbatch[k]

    labels = {k: torch.stack(v, 0) if len(v) else torch.empty(0) for k, v in labels.items()}
    if self.task != "obb" and labels["bboxes"].shape[0]:
        labels["bboxes"] = xyxy2xywh(labels["bboxes"])
    (save_dir / "visualizations").mkdir(parents=True, exist_ok=True)
    plot_images(
        labels,
        img.repeat(4, 1, 1, 1),
        paths=["Ground Truth", "False Positives", "True Positives", "False Negatives"],
        fname=save_dir / "visualizations" / Path(im_file).name,
        names=self.names,
        max_subplots=4,
        conf_thres=0.001,
        show_labels=show_labels,
        show_conf=show_conf,
    )

def print(self):
    """Print the confusion matrix to the console."""
    for i in range(self.matrix.shape[0]):
        LOGGER.info(" ".join(map(str, self.matrix[i])))

def process_batch(
    self,
    detections: dict[str, torch.Tensor],
    batch: dict[str, Any],
    conf: float = 0.25,
    iou_thres: float = 0.45,
) -> None:
    """Update confusion matrix for object detection task.

    Args:
        detections (dict[str, torch.Tensor]): Dictionary containing detected bounding boxes and their associated
            information. Should contain 'cls', 'conf', and 'bboxes' keys, where 'bboxes' can be Array[N, 4] for
            regular boxes or Array[N, 5] for OBB with angle.
        batch (dict[str, Any]): Batch dictionary containing ground truth data with 'bboxes' (Array[M, 4]| Array[M,
            5]) and 'cls' (Array[M]) keys, where M is the number of ground truth objects.
        conf (float, optional): Confidence threshold for detections.
        iou_thres (float, optional): IoU threshold for matching detections to ground truth.
    """
    gt_cls, gt_bboxes = batch["cls"], batch["bboxes"]
    if self.matches is not None:  # only if visualization is enabled
        self.matches = {k: defaultdict(list) for k in {"TP", "FP", "FN", "GT"}}
        for i in range(gt_cls.shape[0]):
            self._append_matches("GT", batch, i)  # store GT
    is_obb = gt_bboxes.shape[1] == 5  # check if boxes contains angle for OBB
    conf = 0.25 if conf in {None, 0.01 if is_obb else 0.001} else conf  # apply 0.25 if default val conf is passed
    no_pred = detections["cls"].shape[0] == 0
    if gt_cls.shape[0] == 0:  # Check if labels is empty
        if not no_pred:
            detections = {k: detections[k][detections["conf"] > conf] for k in detections}
            detection_classes = detections["cls"].int().tolist()
            for i, dc in enumerate(detection_classes):
                self.matrix[dc, self.nc] += 1  # FP
                self._append_matches("FP", detections, i)
        return
    if no_pred:
        gt_classes = gt_cls.int().tolist()
        for i, gc in enumerate(gt_classes):
            self.matrix[self.nc, gc] += 1  # FN
            self._append_matches("FN", batch, i)
        return

    detections = {k: detections[k][detections["conf"] > conf] for k in detections}
    gt_classes = gt_cls.int().tolist()
    detection_classes = detections["cls"].int().tolist()
    bboxes = detections["bboxes"]
    iou = batch_probiou(gt_bboxes, bboxes) if is_obb else box_iou(gt_bboxes, bboxes)

    x = torch.where(iou > iou_thres)
    if x[0].shape[0]:
        matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
        if x[0].shape[0] > 1:
            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]]
    else:
        matches = np.zeros((0, 3))

    n = matches.shape[0] > 0
    m0, m1, _ = matches.transpose().astype(int)
    for i, gc in enumerate(gt_classes):
        j = m0 == i
        if n and sum(j) == 1:
            dc = detection_classes[m1[j].item()]
            self.matrix[dc, gc] += 1  # TP if class is correct else both an FP and an FN
            if dc == gc:
                self._append_matches("TP", detections, m1[j].item())
            else:
                self._append_matches("FP", detections, m1[j].item())
                self._append_matches("FN", batch, i)
        else:
            self.matrix[self.nc, gc] += 1  # FN
            self._append_matches("FN", batch, i)

    for i, dc in enumerate(detection_classes):
        if not any(m1 == i):
            self.matrix[dc, self.nc] += 1  # FP
            self._append_matches("FP", detections, i)

def process_cls_preds(self, preds: list[torch.Tensor], targets: list[torch.Tensor]) -> None:
    """Update confusion matrix for classification task.

    Args:
        preds (list[torch.Tensor]): Predicted class labels.
        targets (list[torch.Tensor]): Ground truth class labels.
    """
    preds, targets = torch.cat(preds)[:, 0], torch.cat(targets)
    for p, t in zip(preds.cpu().numpy(), targets.cpu().numpy()):
        self.matrix[p][t] += 1

def summary(self, normalize: bool = False, decimals: int = 5) -> list[dict[str, float]]:
    """Generate a summarized representation of the confusion matrix as a list of dictionaries, with optional
    normalization. This is useful for exporting the matrix to various formats such as CSV, XML, HTML, JSON,
    or SQL.

    Args:
        normalize (bool): Whether to normalize the confusion matrix values.
        decimals (int): Number of decimal places to round the output values to.

    Returns:
        (list[dict[str, float]]): A list of dictionaries, each representing one predicted class with corresponding
            values for all actual classes.

    Examples:
        >>> results = model.val(data="coco8.yaml", plots=True)
        >>> cm_dict = results.confusion_matrix.summary(normalize=True, decimals=5)
        >>> print(cm_dict)
    """
    import re

    names = (
        list(self.names.values())
        if self.task in {"classify", "semantic"}
        else [*list(self.names.values()), "background"]
    )
    clean_names, seen = [], set()
    for name in names:
        clean_name = re.sub(r"[^a-zA-Z0-9_]", "_", name)
        original_clean = clean_name
        counter = 1
        while clean_name.lower() in seen:
            clean_name = f"{original_clean}_{counter}"
            counter += 1
        seen.add(clean_name.lower())
        clean_names.append(clean_name)
    array = (self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1e-9) if normalize else 1)).round(decimals)
    return [
        dict({"Predicted": clean_names[i]}, **{clean_names[j]: array[i, j] for j in range(len(clean_names))})
        for i in range(len(clean_names))
    ]

def tp_fp(self) -> tuple[np.ndarray, np.ndarray]:
    """Return true positives and false positives.

    Returns:
        tp (np.ndarray): True positives.
        fp (np.ndarray): False positives.
    """
    tp = self.matrix.diagonal()  # true positives
    fp = self.matrix.sum(1) - tp  # false positives
    # fn = self.matrix.sum(0) - tp  # false negatives (missed detections)
    return (tp, fp) if self.task == "classify" else (tp[:-1], fp[:-1])  # remove background class if task=detect

class Metric(SimpleClass):
    """Class for computing evaluation metrics for Ultralytics YOLO models.

    Attributes:
        p (list): Precision for each class. Shape: (nc,).
        r (list): Recall for each class. Shape: (nc,).
        f1 (list): F1 score for each class. Shape: (nc,).
        all_ap (list): AP scores for all classes and all IoU thresholds. Shape: (nc, 10).
        ap_class_index (list): Index of class for each AP score. Shape: (nc,).
        nc (int): Number of classes.

    Methods:
        ap50: AP at IoU threshold of 0.5 for all classes.
        ap: AP at IoU thresholds from 0.5 to 0.95 for all classes.
        mp: Mean precision of all classes.
        mr: Mean recall of all classes.
        map50: Mean AP at IoU threshold of 0.5 for all classes.
        map75: Mean AP at IoU threshold of 0.75 for all classes.
        map: Mean AP at IoU thresholds from 0.5 to 0.95 for all classes.
        mean_results: Mean of results, returns mp, mr, map50, map.
        class_result: Class-aware result, returns p[i], r[i], ap50[i], ap[i].
        maps: mAP of each class.
        fitness: Model fitness as a weighted combination of metrics.
        update: Update metric attributes with new evaluation results.
        curves: Provides a list of curves for accessing specific metrics like precision, recall, F1, etc.
        curves_results: Provide a list of results for accessing specific metrics like precision, recall, F1, etc.
    """

    def __init__(self) -> None:
        """Initialize a Metric instance for computing evaluation metrics for the YOLO model."""
        self.p = []  # (nc, )
        self.r = []  # (nc, )
        self.f1 = []  # (nc, )
        self.all_ap = []  # (nc, 10)
        self.ap_class_index = []  # (nc, )
        self.nc = 0
        self.image_metrics = {}

@property
def ap50(self) -> np.ndarray | list:
    """Return the Average Precision (AP) at an IoU threshold of 0.5 for all classes.

    Returns:
        (np.ndarray | list): Array of shape (nc,) with AP50 values per class, or an empty list if not available.
    """
    return self.all_ap[:, 0] if len(self.all_ap) else []

@property
def ap(self) -> np.ndarray | list:
    """Return the Average Precision (AP) at an IoU threshold of 0.5-0.95 for all classes.

    Returns:
        (np.ndarray | list): Array of shape (nc,) with AP50-95 values per class, or an empty list if not available.
    """
    return self.all_ap.mean(1) if len(self.all_ap) else []

@property
def mp(self) -> float:
    """Return the Mean Precision of all classes.

    Returns:
        (float): The mean precision of all classes.
    """
    return self.p.mean() if len(self.p) else 0.0

@property
def mr(self) -> float:
    """Return the Mean Recall of all classes.

    Returns:
        (float): The mean recall of all classes.
    """
    return self.r.mean() if len(self.r) else 0.0

@property
def map50(self) -> float:
    """Return the mean Average Precision (mAP) at an IoU threshold of 0.5.

    Returns:
        (float): The mAP at an IoU threshold of 0.5.
    """
    return self.all_ap[:, 0].mean() if len(self.all_ap) else 0.0

@property
def map75(self) -> float:
    """Return the mean Average Precision (mAP) at an IoU threshold of 0.75.

    Returns:
        (float): The mAP at an IoU threshold of 0.75.
    """
    return self.all_ap[:, 5].mean() if len(self.all_ap) else 0.0

@property
def map(self) -> float:
    """Return the mean Average Precision (mAP) over IoU thresholds of 0.5 - 0.95 in steps of 0.05.

    Returns:
        (float): The mAP over IoU thresholds of 0.5 - 0.95 in steps of 0.05.
    """
    return self.all_ap.mean() if len(self.all_ap) else 0.0

@property
def maps(self) -> np.ndarray:
    """Return mAP of each class."""
    maps = np.zeros(self.nc) + self.map
    for i, c in enumerate(self.ap_class_index):
        maps[c] = self.ap[i]
    return maps

@property
def curves(self) -> list:
    """Return a list of curves for accessing specific metrics curves."""
    return []

@property
def curves_results(self) -> list[list]:
    """Return a list of curves results for accessing specific metrics curves."""
    return [
        [self.px, self.prec_values, "Recall", "Precision"],
        [self.px, self.f1_curve, "Confidence", "F1"],
        [self.px, self.p_curve, "Confidence", "Precision"],
        [self.px, self.r_curve, "Confidence", "Recall"],
    ]

def class_result(self, i: int) -> tuple[float, float, float, float]:
    """Return class-aware result, p[i], r[i], ap50[i], ap[i]."""
    return self.p[i], self.r[i], self.ap50[i], self.ap[i]

def clear_image_metrics(self) -> None:
    """Clear stored per-image metrics from the current validation run."""
    self.image_metrics.clear()

def fitness(self) -> float:
    """Return model fitness as a weighted combination of metrics."""
    w = [0.0, 0.0, 0.0, 1.0]  # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
    return float((np.nan_to_num(np.array(self.mean_results())) * w).sum())

def mean_results(self) -> list[float]:
    """Return mean of results, mp, mr, map50, map."""
    return [self.mp, self.mr, self.map50, self.map]

def update(self, results: tuple):
    """Update the evaluation metrics with a new set of results.

    Args:
        results (tuple): A tuple containing evaluation metrics:
            - p (list): Precision for each class.
            - r (list): Recall for each class.
            - f1 (list): F1 score for each class.
            - all_ap (list): AP scores for all classes and all IoU thresholds.
            - ap_class_index (list): Index of class for each AP score.
            - p_curve (list): Precision curve for each class.
            - r_curve (list): Recall curve for each class.
            - f1_curve (list): F1 curve for each class.
            - px (list): X values for the curves.
            - prec_values (list): Precision values for each class.
    """
    (
        self.p,
        self.r,
        self.f1,
        self.all_ap,
        self.ap_class_index,
        self.p_curve,
        self.r_curve,
        self.f1_curve,
        self.px,
        self.prec_values,
    ) = results

def update_image_metrics(self, tp: np.ndarray, target_cls: np.ndarray, pred_cls: np.ndarray, im_name: str) -> None:
    """Update per-image precision, recall, F1, TP, FP, and FN at IoU threshold 0.5.

    Args:
        tp (np.ndarray): True positive array of shape (num_preds, num_iou_thresholds), where the first column (IoU
            >= 0.5) is used.
        target_cls (np.ndarray): Ground truth class labels for the image.
        pred_cls (np.ndarray): Predicted class labels for the image.
        im_name (str): The image filename used as the per-image key.
    """
    # Use the default IoU=0.5 column to match the validator's image-level matching policy.
    tp = int(tp[:, 0].sum())
    num_preds = pred_cls.shape[0]
    num_targets = target_cls.shape[0]
    fp = num_preds - tp
    fn = num_targets - tp
    if num_preds == 0 and num_targets == 0:
        # Empty-GT image with no predictions is a trivially correct call, so report a perfect score rather than
        # zeroing out P/R/F1 by the standard 0/0 fallback below.
        precision = recall = f1 = 1.0
    else:
        precision = tp / num_preds if num_preds else 0.0
        recall = tp / num_targets if num_targets else 0.0
        denom = precision + recall
        f1 = 2 * precision * recall / denom if denom else 0.0
    self.image_metrics[im_name] = {
        "precision": float(precision),
        "recall": float(recall),
        "f1": float(f1),
        "tp": int(tp),
        "fp": int(fp),
        "fn": int(fn),
    }

class DetMetrics(SimpleClass, DataExportMixin):
    """Utility class for computing detection metrics such as precision, recall, and mean average precision (mAP).

    Attributes:
        names (dict[int, str]): A dictionary of class names.
        box (Metric): An instance of the Metric class for storing detection results.
        speed (dict[str, float]): A dictionary for storing execution times of different parts of the detection process.
        stats (dict[str, list]): A dictionary containing lists for true positives, confidence scores, predicted classes,
            target classes, and target images.
        nt_per_class: Number of targets per class.
        nt_per_image: Number of targets per image.

    Methods:
        update_stats: Update statistics by appending new values to existing stat collections.
        process: Process predicted results for object detection and update metrics.
        clear_stats: Clear the stored statistics.
        keys: Return a list of keys for accessing specific metrics.
        mean_results: Calculate mean of detected objects & return precision, recall, mAP50, and mAP50-95.
        class_result: Return the result of evaluating the performance of an object detection model on a specific class.
        maps: Return mean Average Precision (mAP) scores per class.
        fitness: Return the fitness of box object.
        ap_class_index: Return the average precision index per class.
        results_dict: Return dictionary of computed performance metrics and statistics.
        curves: Return a list of curves for accessing specific metrics curves.
        curves_results: Return a list of computed performance metrics and statistics.
        summary: Generate a summarized representation of per-class detection metrics as a list of dictionaries.
    """

    def __init__(self, names: dict[int, str] = {}) -> None:
        """Initialize a DetMetrics instance with class names.

        Args:
            names (dict[int, str], optional): Dictionary of class names.
        """
        self.names = names
        self.box = Metric()
        self.speed = {"preprocess": 0.0, "inference": 0.0, "loss": 0.0, "postprocess": 0.0}
        self.stats = dict(tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
        self.nt_per_class = None
        self.nt_per_image = None

@property
def keys(self) -> list[str]:
    """Return a list of keys for accessing specific metrics."""
    return ["metrics/precision(B)", "metrics/recall(B)", "metrics/mAP50(B)", "metrics/mAP50-95(B)"]

@property
def maps(self) -> np.ndarray:
    """Return mean Average Precision (mAP) scores per class."""
    return self.box.maps

@property
def fitness(self) -> float:
    """Return the fitness of box object."""
    return self.box.fitness()

@property
def ap_class_index(self) -> list:
    """Return the average precision index per class."""
    return self.box.ap_class_index

@property
def results_dict(self) -> dict[str, float]:
    """Return dictionary of computed performance metrics and statistics."""
    keys = [*self.keys, "fitness"]
    values = ((float(x) if hasattr(x, "item") else x) for x in ([*self.mean_results(), self.fitness]))
    return dict(zip(keys, values))

@property
def curves(self) -> list[str]:
    """Return a list of curves for accessing specific metrics curves."""
    return ["Precision-Recall(B)", "F1-Confidence(B)", "Precision-Confidence(B)", "Recall-Confidence(B)"]

@property
def curves_results(self) -> list[list]:
    """Return a list of computed performance metrics and statistics."""
    return self.box.curves_results

def class_result(self, i: int) -> tuple[float, float, float, float]:
    """Return the result of evaluating the performance of an object detection model on a specific class."""
    return self.box.class_result(i)

def clear_image_metrics(self) -> None:
    """Clear stored per-image metrics."""
    self.box.clear_image_metrics()

def clear_stats(self):
    """Clear the stored statistics."""
    for v in self.stats.values():
        v.clear()

def mean_results(self) -> list[float]:
    """Calculate mean of detected objects & return precision, recall, mAP50, and mAP50-95."""
    return self.box.mean_results()

def process(self, save_dir: Path = Path("."), plot: bool = False, on_plot=None) -> dict[str, np.ndarray]:
    """Process predicted results for object detection and update metrics.

    Args:
        save_dir (Path): Directory to save plots. Defaults to Path(".").
        plot (bool): Whether to plot precision-recall curves. Defaults to False.
        on_plot (callable, optional): Function to call after plots are generated. Defaults to None.

    Returns:
        (dict[str, np.ndarray]): Dictionary containing concatenated statistics arrays.
    """
    stats = {k: np.concatenate(v, 0) for k, v in self.stats.items()}  # to numpy
    if not stats:
        return stats
    results = ap_per_class(
        stats["tp"],
        stats["conf"],
        stats["pred_cls"],
        stats["target_cls"],
        plot=plot,
        save_dir=save_dir,
        names=self.names,
        on_plot=on_plot,
        prefix="Box",
    )[2:]
    self.box.nc = len(self.names)
    self.box.update(results)
    self.nt_per_class = np.bincount(stats["target_cls"].astype(int), minlength=len(self.names))
    self.nt_per_image = np.bincount(stats["target_img"].astype(int), minlength=len(self.names))
    return stats

def summary(self, normalize: bool = True, decimals: int = 5) -> list[dict[str, Any]]:
    """Generate a summarized representation of per-class detection metrics as a list of dictionaries. Includes
    shared scalar metrics (mAP, mAP50, mAP75) alongside precision, recall, and F1-score for each class.

    Args:
        normalize (bool): For Detect metrics, everything is normalized by default [0-1].
        decimals (int): Number of decimal places to round the metrics values to.

    Returns:
        (list[dict[str, Any]]): A list of dictionaries, each representing one class with corresponding metric
            values.

    Examples:
       >>> results = model.val(data="coco8.yaml")
       >>> detection_summary = results.summary()
       >>> print(detection_summary)
    """
    per_class = {
        "Box-P": self.box.p,
        "Box-R": self.box.r,
        "Box-F1": self.box.f1,
    }
    return [
        {
            "Class": self.names[self.ap_class_index[i]],
            "Images": self.nt_per_image[self.ap_class_index[i]],
            "Instances": self.nt_per_class[self.ap_class_index[i]],
            **{k: round(v[i], decimals) for k, v in per_class.items()},
            "mAP50": round(self.class_result(i)[2], decimals),
            "mAP50-95": round(self.class_result(i)[3], decimals),
        }
        for i in range(len(per_class["Box-P"]))
    ]

def update_stats(self, stat: dict[str, Any]) -> None:
    """Update statistics by appending new values to existing stat collections.

    Args:
        stat (dict[str, Any]): Dictionary containing new statistical values to append. Keys should match existing
            keys in self.stats.
    """
    for k in self.stats.keys():
        self.stats[k].append(stat[k])
    self.box.update_image_metrics(stat["tp"], stat["target_cls"], stat["pred_cls"], stat["im_name"])

class SegmentMetrics(DetMetrics):
    """Calculate and aggregate detection and segmentation metrics over a given set of classes.

    Attributes:
        names (dict[int, str]): Dictionary of class names.
        box (Metric): An instance of the Metric class for storing detection results.
        seg (Metric): An instance of the Metric class to calculate mask segmentation metrics.
        speed (dict[str, float]): A dictionary for storing execution times of different parts of the detection process.
        stats (dict[str, list]): A dictionary containing lists for true positives, confidence scores, predicted classes,
            target classes, and target images.
        nt_per_class: Number of targets per class.
        nt_per_image: Number of targets per image.

    Methods:
        process: Process the detection and segmentation metrics over the given set of predictions.
        keys: Return a list of keys for accessing metrics.
        mean_results: Return the mean metrics for bounding box and segmentation results.
        class_result: Return classification results for a specified class index.
        maps: Return mAP scores for object detection and segmentation models.
        fitness: Return the fitness score for both segmentation and bounding box models.
        curves: Return a list of curves for accessing specific metrics curves.
        curves_results: Provide a list of computed performance metrics and statistics.
        summary: Generate a summarized representation of per-class segmentation metrics as a list of dictionaries.
    """

    def __init__(self, names: dict[int, str] = {}) -> None:
        """Initialize a SegmentMetrics instance with class names.

        Args:
            names (dict[int, str], optional): Dictionary of class names.
        """
        DetMetrics.__init__(self, names)
        self.seg = Metric()
        self.stats["tp_m"] = []  # add additional stats for masks

@property
def keys(self) -> list[str]:
    """Return a list of keys for accessing metrics."""
    return [
        *DetMetrics.keys.fget(self),
        "metrics/precision(M)",
        "metrics/recall(M)",
        "metrics/mAP50(M)",
        "metrics/mAP50-95(M)",
    ]

@property
def maps(self) -> np.ndarray:
    """Return mAP scores for object detection and segmentation models."""
    return DetMetrics.maps.fget(self) + self.seg.maps

@property
def fitness(self) -> float:
    """Return the fitness score for both segmentation and bounding box models."""
    return self.seg.fitness() + DetMetrics.fitness.fget(self)

@property
def curves(self) -> list[str]:
    """Return a list of curves for accessing specific metrics curves."""
    return [
        *DetMetrics.curves.fget(self),
        "Precision-Recall(M)",
        "F1-Confidence(M)",
        "Precision-Confidence(M)",
        "Recall-Confidence(M)",
    ]

@property
def curves_results(self) -> list[list]:
    """Return a list of computed performance metrics and statistics."""
    return DetMetrics.curves_results.fget(self) + self.seg.curves_results

def class_result(self, i: int) -> list[float]:
    """Return classification results for a specified class index."""
    return DetMetrics.class_result(self, i) + self.seg.class_result(i)

def clear_image_metrics(self) -> None:
    """Clear stored per-image metrics."""
    super().clear_image_metrics()
    self.seg.clear_image_metrics()

def mean_results(self) -> list[float]:
    """Return the mean metrics for bounding box and segmentation results."""
    return DetMetrics.mean_results(self) + self.seg.mean_results()

def process(self, save_dir: Path = Path("."), plot: bool = False, on_plot=None) -> dict[str, np.ndarray]:
    """Process the detection and segmentation metrics over the given set of predictions.

    Args:
        save_dir (Path): Directory to save plots. Defaults to Path(".").
        plot (bool): Whether to plot precision-recall curves. Defaults to False.
        on_plot (callable, optional): Function to call after plots are generated. Defaults to None.

    Returns:
        (dict[str, np.ndarray]): Dictionary containing concatenated statistics arrays.
    """
    stats = DetMetrics.process(self, save_dir, plot, on_plot=on_plot)  # process box stats
    results_mask = ap_per_class(
        stats["tp_m"],
        stats["conf"],
        stats["pred_cls"],
        stats["target_cls"],
        plot=plot,
        on_plot=on_plot,
        save_dir=save_dir,
        names=self.names,
        prefix="Mask",
    )[2:]
    self.seg.nc = len(self.names)
    self.seg.update(results_mask)
    return stats

def summary(self, normalize: bool = True, decimals: int = 5) -> list[dict[str, Any]]:
    """Generate a summarized representation of per-class segmentation metrics as a list of dictionaries. Includes
    both box and mask scalar metrics (mAP, mAP50, mAP75) alongside precision, recall, and F1-score for
    each class.

    Args:
        normalize (bool): For Segment metrics, everything is normalized by default [0-1].
        decimals (int): Number of decimal places to round the metrics values to.

    Returns:
        (list[dict[str, Any]]): A list of dictionaries, each representing one class with corresponding metric
            values.

    Examples:
        >>> results = model.val(data="coco8-seg.yaml")
        >>> seg_summary = results.summary(decimals=4)
        >>> print(seg_summary)
    """
    per_class = {
        "Mask-P": self.seg.p,
        "Mask-R": self.seg.r,
        "Mask-F1": self.seg.f1,
    }
    summary = DetMetrics.summary(self, normalize, decimals)  # get box summary
    for i, s in enumerate(summary):
        s.update({**{k: round(v[i], decimals) for k, v in per_class.items()}})
    return summary

def update_stats(self, stat: dict[str, Any]) -> None:
    """Update statistics by appending new values to existing stat collections.

    Args:
        stat (dict[str, Any]): Dictionary containing new statistical values to append. Keys should match existing
            keys in self.stats.
    """
    super().update_stats(stat)  # update box stats
    self.seg.update_image_metrics(stat["tp_m"], stat["target_cls"], stat["pred_cls"], stat["im_name"])

class PoseMetrics(DetMetrics):
    """Calculate and aggregate detection and pose metrics over a given set of classes.

    Attributes:
        names (dict[int, str]): Dictionary of class names.
        pose (Metric): An instance of the Metric class to calculate pose metrics.
        box (Metric): An instance of the Metric class for storing detection results.
        speed (dict[str, float]): A dictionary for storing execution times of different parts of the detection process.
        stats (dict[str, list]): A dictionary containing lists for true positives, confidence scores, predicted classes,
            target classes, and target images.
        nt_per_class: Number of targets per class.
        nt_per_image: Number of targets per image.

    Methods:
        process: Process the detection and pose metrics over the given set of predictions.
        keys: Return a list of keys for accessing metrics.
        mean_results: Return the mean results of box and pose.
        class_result: Return the class-wise detection results for a specific class i.
        maps: Return the mean average precision (mAP) per class for both box and pose detections.
        fitness: Return combined fitness score for pose and box detection.
        curves: Return a list of curves for accessing specific metrics curves.
        curves_results: Provide a list of computed performance metrics and statistics.
        summary: Generate a summarized representation of per-class pose metrics as a list of dictionaries.
    """

    def __init__(self, names: dict[int, str] = {}) -> None:
        """Initialize the PoseMetrics class with class names.

        Args:
            names (dict[int, str], optional): Dictionary of class names.
        """
        super().__init__(names)
        self.pose = Metric()
        self.stats["tp_p"] = []  # add additional stats for pose

@property
def keys(self) -> list[str]:
    """Return a list of evaluation metric keys."""
    return [
        *DetMetrics.keys.fget(self),
        "metrics/precision(P)",
        "metrics/recall(P)",
        "metrics/mAP50(P)",
        "metrics/mAP50-95(P)",
    ]

@property
def maps(self) -> np.ndarray:
    """Return the mean average precision (mAP) per class for both box and pose detections."""
    return DetMetrics.maps.fget(self) + self.pose.maps

@property
def fitness(self) -> float:
    """Return combined fitness score for pose and box detection."""
    return self.pose.fitness() + DetMetrics.fitness.fget(self)

@property
def curves(self) -> list[str]:
    """Return a list of curves for accessing specific metrics curves."""
    return [
        *DetMetrics.curves.fget(self),
        "Precision-Recall(B)",
        "F1-Confidence(B)",
        "Precision-Confidence(B)",
        "Recall-Confidence(B)",
        "Precision-Recall(P)",
        "F1-Confidence(P)",
        "Precision-Confidence(P)",
        "Recall-Confidence(P)",
    ]

@property
def curves_results(self) -> list[list]:
    """Return a list of computed performance metrics and statistics."""
    return DetMetrics.curves_results.fget(self) + self.pose.curves_results

def class_result(self, i: int) -> list[float]:
    """Return the class-wise detection results for a specific class i."""
    return DetMetrics.class_result(self, i) + self.pose.class_result(i)

def clear_image_metrics(self) -> None:
    """Clear stored per-image metrics."""
    super().clear_image_metrics()
    self.pose.clear_image_metrics()

def mean_results(self) -> list[float]:
    """Return the mean results of box and pose."""
    return DetMetrics.mean_results(self) + self.pose.mean_results()

def process(self, save_dir: Path = Path("."), plot: bool = False, on_plot=None) -> dict[str, np.ndarray]:
    """Process the detection and pose metrics over the given set of predictions.

    Args:
        save_dir (Path): Directory to save plots. Defaults to Path(".").
        plot (bool): Whether to plot precision-recall curves. Defaults to False.
        on_plot (callable, optional): Function to call after plots are generated.

    Returns:
        (dict[str, np.ndarray]): Dictionary containing concatenated statistics arrays.
    """
    stats = DetMetrics.process(self, save_dir, plot, on_plot=on_plot)  # process box stats
    results_pose = ap_per_class(
        stats["tp_p"],
        stats["conf"],
        stats["pred_cls"],
        stats["target_cls"],
        plot=plot,
        on_plot=on_plot,
        save_dir=save_dir,
        names=self.names,
        prefix="Pose",
    )[2:]
    self.pose.nc = len(self.names)
    self.pose.update(results_pose)
    return stats

def summary(self, normalize: bool = True, decimals: int = 5) -> list[dict[str, Any]]:
    """Generate a summarized representation of per-class pose metrics as a list of dictionaries. Includes both box
    and pose scalar metrics (mAP, mAP50, mAP75) alongside precision, recall, and F1-score for each class.

    Args:
        normalize (bool): For Pose metrics, everything is normalized by default [0-1].
        decimals (int): Number of decimal places to round the metrics values to.

    Returns:
        (list[dict[str, Any]]): A list of dictionaries, each representing one class with corresponding metric
            values.

    Examples:
        >>> results = model.val(data="coco8-pose.yaml")
        >>> pose_summary = results.summary(decimals=4)
        >>> print(pose_summary)
    """
    per_class = {
        "Pose-P": self.pose.p,
        "Pose-R": self.pose.r,
        "Pose-F1": self.pose.f1,
    }
    summary = DetMetrics.summary(self, normalize, decimals)  # get box summary
    for i, s in enumerate(summary):
        s.update({**{k: round(v[i], decimals) for k, v in per_class.items()}})
    return summary

def update_stats(self, stat: dict[str, Any]) -> None:
    """Update statistics by appending new values to existing stat collections.

    Args:
        stat (dict[str, Any]): Dictionary containing new statistical values to append. Keys should match existing
            keys in self.stats.
    """
    super().update_stats(stat)  # update box stats
    self.pose.update_image_metrics(stat["tp_p"], stat["target_cls"], stat["pred_cls"], stat["im_name"])

class ClassifyMetrics(SimpleClass, DataExportMixin):
    """Class for computing classification metrics including top-1 and top-5 accuracy.

    Attributes:
        top1 (float): The top-1 accuracy.
        top5 (float): The top-5 accuracy.
        speed (dict[str, float]): A dictionary containing the time taken for each step in the pipeline.

    Methods:
        process: Process target classes and predicted classes to compute metrics.
        fitness: Return mean of top-1 and top-5 accuracies as fitness score.
        results_dict: Return a dictionary with model's performance metrics and fitness score.
        keys: Return a list of keys for the results_dict property.
        curves: Return a list of curves for accessing specific metrics curves.
        curves_results: Provide a list of computed performance metrics and statistics.
        summary: Generate a single-row summary of classification metrics (Top-1 and Top-5 accuracy).
    """

    def __init__(self) -> None:
        """Initialize a ClassifyMetrics instance."""
        self.top1 = 0
        self.top5 = 0
        self.speed = {"preprocess": 0.0, "inference": 0.0, "loss": 0.0, "postprocess": 0.0}

@property
def fitness(self) -> float:
    """Return mean of top-1 and top-5 accuracies as fitness score."""
    return (self.top1 + self.top5) / 2

@property
def results_dict(self) -> dict[str, float]:
    """Return a dictionary with model's performance metrics and fitness score."""
    return dict(zip([*self.keys, "fitness"], [self.top1, self.top5, self.fitness]))

@property
def keys(self) -> list[str]:
    """Return a list of keys for the results_dict property."""
    return ["metrics/accuracy_top1", "metrics/accuracy_top5"]

@property
def curves(self) -> list:
    """Return a list of curves for accessing specific metrics curves."""
    return []

@property
def curves_results(self) -> list:
    """Return a list of curves results for accessing specific metrics curves."""
    return []

def process(self, targets: torch.Tensor, pred: torch.Tensor):
    """Process target classes and predicted classes to compute metrics.

    Args:
        targets (torch.Tensor): Target classes.
        pred (torch.Tensor): Predicted classes.
    """
    pred, targets = torch.cat(pred), torch.cat(targets)
    correct = (targets[:, None] == pred).float()
    acc = torch.stack((correct[:, 0], correct.max(1).values), dim=1)  # (top1, top5) accuracy
    self.top1, self.top5 = acc.mean(0).tolist()

def summary(self, normalize: bool = True, decimals: int = 5) -> list[dict[str, float]]:
    """Generate a single-row summary of classification metrics (Top-1 and Top-5 accuracy).

    Args:
        normalize (bool): For Classify metrics, everything is normalized by default [0-1].
        decimals (int): Number of decimal places to round the metrics values to.

    Returns:
        (list[dict[str, float]]): A list with one dictionary containing Top-1 and Top-5 classification accuracy.

    Examples:
        >>> results = model.val(data="imagenet10")
        >>> classify_summary = results.summary(decimals=4)
        >>> print(classify_summary)
    """
    return [{"top1_acc": round(self.top1, decimals), "top5_acc": round(self.top5, decimals)}]

class OBBMetrics(DetMetrics):
    """Metrics for evaluating oriented bounding box (OBB) detection.

    Attributes:
        names (dict[int, str]): Dictionary of class names.
        box (Metric): An instance of the Metric class for storing detection results.
        speed (dict[str, float]): A dictionary for storing execution times of different parts of the detection process.
        stats (dict[str, list]): A dictionary containing lists for true positives, confidence scores, predicted classes,
            target classes, and target images.
        nt_per_class: Number of targets per class.
        nt_per_image: Number of targets per image.

    References:
        https://arxiv.org/pdf/2106.06072.pdf
    """

    def __init__(self, names: dict[int, str] = {}) -> None:
        """Initialize an OBBMetrics instance with class names.

        Args:
            names (dict[int, str], optional): Dictionary of class names.
        """
        DetMetrics.__init__(self, names)

class SemanticMetrics(SimpleClass, DataExportMixin):
    """Metrics for semantic segmentation, including mIoU, pixel accuracy, and per-class IoU.

    Attributes:
        names (dict): Class names mapping.
        nc (int): Number of classes.
        cm_nc (int): Confusion matrix side length (2 for binary segmentation, else nc).
        device (torch.device | None): Device used for confusion matrix accumulation.
        matrix (torch.Tensor | None): Accumulated confusion matrix of shape (cm_nc, cm_nc).
        speed (dict): Processing speed statistics.
        nt_per_image (np.ndarray): Number of images containing each class.
        nt_per_class (np.ndarray): Number of pixels per class.
        _miou (float): Cached mean IoU.
        _pixel_accuracy (float): Cached pixel accuracy.
        _per_class_iou (np.ndarray): Cached per-class IoU values.
        _per_class_pixel_acc (np.ndarray): Cached per-class pixel accuracy.
    """

    def __init__(self, names: dict[int, str] | None = None) -> None:
        """Initialize semantic segmentation metrics.

        Args:
            names (dict, optional): Dictionary mapping class indices to names.
        """
        self.names = names or {}
        self.nc = len(self.names)
        self.cm_nc = 2 if self.nc == 1 else self.nc
        self.matrix = None
        self.speed = {"preprocess": 0.0, "inference": 0.0, "loss": 0.0, "postprocess": 0.0}
        self.nt_per_image = np.zeros(self.nc, dtype=np.int32)
        self._miou = 0.0
        self._pixel_accuracy = 0.0
        self._per_class_iou = np.zeros(self.nc, dtype=np.float32)
        self._per_class_pixel_acc = np.zeros(self.nc, dtype=np.float32)
        self.nt_per_class = np.zeros(self.nc, dtype=np.int32)

@property
def miou(self):
    """Return mean IoU (foreground IoU only for binary segmentation)."""
    return self._miou

@property
def pixel_accuracy(self):
    """Return overall pixel accuracy."""
    return self._pixel_accuracy

@property
def per_class_iou(self):
    """Return per-class IoU values (foreground IoU only for binary segmentation)."""
    return self._per_class_iou

@property
def per_class_pixel_accuracy(self):
    """Return per-class pixel accuracy (diagonal / row sum for each class)."""
    return self._per_class_pixel_acc

@property
def fitness(self):
    """Return model fitness as mean IoU."""
    return self.miou

@property
def keys(self):
    """Return metric keys for logging."""
    return ["metrics/mIoU", "metrics/pixel_acc"]

@property
def ap_class_index(self):
    """Return the class index list for per-class results."""
    return list(range(self.nc))

@property
def results_dict(self):
    """Return results dictionary."""
    return dict(zip([*self.keys, "fitness"], [*self.mean_results(), self.fitness]))

@property
def curves(self):
    """Return an empty list because semantic segmentation has no PR curves."""
    return []

@property
def curves_results(self):
    """Return empty list (no PR curve results)."""
    return []

@plt_settings()
def _plot_iou_bars(self, save_dir, on_plot):
    """Plot per-class IoU bar chart.

    Args:
        save_dir (Path | str): Directory to save the plot.
        on_plot (callable, optional): Function to call after plot is saved.
    """
    import matplotlib.pyplot as plt

    fig, ax = plt.subplots(1, 1, figsize=(10, 6), tight_layout=True)
    names = list(self.names.values()) if self.names else [str(i) for i in range(self.nc)]
    x = np.arange(self.nc)
    bars = ax.bar(x, self._per_class_iou, color=[list(c / 255.0 for c in colors(i, False)) for i in range(self.nc)])
    ax.set_xlabel("Class")
    ax.set_ylabel("IoU")
    ax.set_title("Per-Class IoU")
    ax.set_ylim(0, 1)
    if 0 < len(names) < 30:
        ax.set_xticks(x)
        ax.set_xticklabels(names, rotation=90, fontsize=10)
    for bar in bars:
        height = bar.get_height()
        ax.text(bar.get_x() + bar.get_width() / 2.0, height, f"{height:.3f}", ha="center", va="bottom", fontsize=8)
    fname = Path(save_dir) / "iou_bar_chart.png"
    plt.savefig(fname, dpi=250)
    plt.close(fig)
    if on_plot:
        on_plot(fname)

def class_result(self, i: int) -> list[float]:
    """Return the result of evaluating the performance on a specific class.

    Args:
        i (int): Class index.

    Returns:
        (list): [IoU, pixel_accuracy] for the specified class.
    """
    if self._per_class_iou is None or len(self._per_class_iou) == 0:
        return [0.0, 0.0]
    return [float(self._per_class_iou[i]), float(self._per_class_pixel_acc[i])]

def clear_stats(self):
    """Clear accumulated statistics."""
    self.matrix = None
    self.nt_per_image.fill(0)

def mean_results(self):
    """Return mean results for logging."""
    return [self.miou, self.pixel_accuracy]

def process(self, save_dir: Path = Path("."), plot: bool = False, on_plot: callable | None = None) -> None:
    """Compute final metrics from accumulated confusion matrix.

    Args:
        save_dir (Path): Directory to save plots. Defaults to Path('.').
        plot (bool): Whether to plot IoU bars and confusion matrix. Defaults to False.
        on_plot (callable, optional): Function to call after plots are generated. Defaults to None.
    """
    if self.matrix is None:
        return

    intersection = torch.diagonal(self.matrix)
    union = self.matrix.sum(1) + self.matrix.sum(0) - intersection
    iou = torch.where(union > 0, intersection / union, torch.zeros_like(intersection, dtype=torch.float32))
    row_sum = self.matrix.sum(1)
    pa = intersection / (row_sum + 1e-10)

    if self.nc == 1:
        self._miou = float(iou[1].item())
        self._per_class_iou = iou[1:].cpu().numpy()
        self._per_class_pixel_acc = pa[1:].cpu().numpy()
        self.nt_per_class = np.array([row_sum[1].item()], dtype=np.int32)
    else:
        self._miou = float(iou.mean().item())
        self._per_class_iou = iou.cpu().numpy()
        self._per_class_pixel_acc = pa.cpu().numpy()
        self.nt_per_class = row_sum[: self.nc].cpu().numpy().astype(np.int32)

    self._pixel_accuracy = float((intersection.sum() / (self.matrix.sum() + 1e-10)).item())

    if plot:
        self._plot_iou_bars(save_dir, on_plot)

def summary(self, normalize: bool = True, decimals: int = 5) -> list[dict]:
    """Generate a per-class summary of semantic segmentation metrics, with global mIoU and pixel accuracy on each
    row.

    Args:
        normalize (bool): For semantic metrics, values are already in [0, 1].
        decimals (int): Number of decimal places to round the metric values to.

    Returns:
        (list[dict]): A list of dictionaries, one per class, with per-class IoU and shared scalars.
    """
    miou = round(self.miou, decimals)
    pixel_acc = round(self.pixel_accuracy, decimals)
    per_class = self.per_class_iou
    names = self.names or {i: str(i) for i in range(len(per_class))}
    return [
        {
            "Class": names.get(i, str(i)),
            "Images": int(self.nt_per_image[i]),
            "Pixels": int(self.nt_per_class[i]),
            "IoU": round(float(per_class[i]), decimals),
            "mIoU": miou,
            "pixel_acc": pixel_acc,
        }
        for i in range(len(per_class))
    ]

def update_stats(self, preds: torch.Tensor, targets: torch.Tensor) -> None:
    """Accumulate confusion matrix from predictions and targets.

    Args:
        preds (torch.Tensor): Predicted class IDs [B, H, W].
        targets (torch.Tensor): Ground truth class IDs [B, H, W].
    """
    if self.matrix is None:
        self.matrix = torch.zeros((self.cm_nc, self.cm_nc), device=preds.device, dtype=torch.float32)

    valid = (targets != 255) & (preds >= 0) & (preds < self.cm_nc) & (targets >= 0) & (targets < self.cm_nc)
    hist = torch.bincount(self.cm_nc * targets[valid] + preds[valid], minlength=self.cm_nc**2).reshape(
        self.cm_nc, self.cm_nc
    )
    self.matrix += hist.to(self.matrix.dtype)

    present = torch.zeros((targets.shape[0], self.cm_nc), dtype=torch.bool, device=targets.device)
    batch_idx = torch.arange(targets.shape[0], device=targets.device).view(-1, 1, 1).expand_as(targets)
    present[batch_idx[valid], targets[valid].long()] = True
    if self.nc == 1:
        self.nt_per_image[0] += int(present[:, 1].sum())
    else:
        self.nt_per_image += present[:, : self.nc].sum(0).cpu().numpy()

def bbox_ioa(box1: np.ndarray, box2: np.ndarray, iou: bool = False, eps: float = 1e-7) -> np.ndarray:
    """Calculate the intersection over box2 area given box1 and box2.

    Args:
        box1 (np.ndarray): A numpy array of shape (N, 4) representing N bounding boxes in x1y1x2y2 format.
        box2 (np.ndarray): A numpy array of shape (M, 4) representing M bounding boxes in x1y1x2y2 format.
        iou (bool, optional): Calculate the standard IoU if True else return inter_area/box2_area.
        eps (float, optional): A small value to avoid division by zero.

    Returns:
        (np.ndarray): A numpy array of shape (N, M) representing the intersection over box2 area.
    """
    # Get the coordinates of bounding boxes
    b1_x1, b1_y1, b1_x2, b1_y2 = box1.T
    b2_x1, b2_y1, b2_x2, b2_y2 = box2.T

    # Intersection area
    inter_area = (np.minimum(b1_x2[:, None], b2_x2) - np.maximum(b1_x1[:, None], b2_x1)).clip(0) * (
        np.minimum(b1_y2[:, None], b2_y2) - np.maximum(b1_y1[:, None], b2_y1)
    ).clip(0)

    # Box2 area
    area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1)
    if iou:
        box1_area = (b1_x2 - b1_x1) * (b1_y2 - b1_y1)
        area = area + box1_area[:, None] - inter_area

    # Intersection over box2 area
    return inter_area / (area + eps)

def box_iou(box1: torch.Tensor, box2: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
    """Calculate intersection-over-union (IoU) of boxes.

    Args:
        box1 (torch.Tensor): A tensor of shape (N, 4) representing N bounding boxes in (x1, y1, x2, y2) format.
        box2 (torch.Tensor): A tensor of shape (M, 4) representing M bounding boxes in (x1, y1, x2, y2) format.
        eps (float, optional): A small value to avoid division by zero.

    Returns:
        (torch.Tensor): An NxM tensor containing the pairwise IoU values for every element in box1 and box2.

    References:
        https://github.com/pytorch/vision/blob/main/torchvision/ops/boxes.py
    """
    # NOTE: Need .float() to get accurate iou values
    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
    (a1, a2), (b1, b2) = box1.float().unsqueeze(1).chunk(2, 2), box2.float().unsqueeze(0).chunk(2, 2)
    inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp_(0).prod(2)

    # IoU = inter / (area1 + area2 - inter)
    return inter / ((a2 - a1).prod(2) + (b2 - b1).prod(2) - inter + eps)

def bbox_iou(
    box1: torch.Tensor,
    box2: torch.Tensor,
    xywh: bool = True,
    GIoU: bool = False,
    DIoU: bool = False,
    CIoU: bool = False,
    eps: float = 1e-7,
) -> torch.Tensor:
    """Calculate the Intersection over Union (IoU) between bounding boxes.

    This function supports various shapes for `box1` and `box2` as long as the last dimension is 4. For instance, you
    may pass tensors shaped like (4,), (N, 4), (B, N, 4), or (B, N, 1, 4). Internally, the code will split the last
    dimension into (x, y, w, h) if `xywh=True`, or (x1, y1, x2, y2) if `xywh=False`.

    Args:
        box1 (torch.Tensor): A tensor representing one or more bounding boxes, with the last dimension being 4.
        box2 (torch.Tensor): A tensor representing one or more bounding boxes, with the last dimension being 4.
        xywh (bool, optional): If True, input boxes are in (x, y, w, h) format. If False, input boxes are in (x1, y1,
            x2, y2) format.
        GIoU (bool, optional): If True, calculate Generalized IoU.
        DIoU (bool, optional): If True, calculate Distance IoU.
        CIoU (bool, optional): If True, calculate Complete IoU.
        eps (float, optional): A small value to avoid division by zero.

    Returns:
        (torch.Tensor): IoU, GIoU, DIoU, or CIoU values depending on the specified flags.
    """
    # Get the coordinates of bounding boxes
    if xywh:  # transform from xywh to xyxy
        (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, -1), box2.chunk(4, -1)
        w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2
        b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_
        b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_
    else:  # x1, y1, x2, y2 = box1
        b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, -1)
        b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, -1)
        w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
        w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps

    # Intersection area
    inter = (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp_(0) * (
        b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)
    ).clamp_(0)

    # Union Area
    union = w1 * h1 + w2 * h2 - inter + eps

    # IoU
    iou = inter / union
    if CIoU or DIoU or GIoU:
        cw = b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1)  # convex (smallest enclosing box) width
        ch = b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1)  # convex height
        if CIoU or DIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
            c2 = cw.pow(2) + ch.pow(2) + eps  # convex diagonal squared
            rho2 = (
                (b2_x1 + b2_x2 - b1_x1 - b1_x2).pow(2) + (b2_y1 + b2_y2 - b1_y1 - b1_y2).pow(2)
            ) / 4  # center dist**2
            if CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
                v = (4 / math.pi**2) * ((w2 / h2).atan() - (w1 / h1).atan()).pow(2)
                with torch.no_grad():
                    alpha = v / (v - iou + (1 + eps))
                return iou - (rho2 / c2 + v * alpha)  # CIoU
            return iou - rho2 / c2  # DIoU
        c_area = cw * ch + eps  # convex area
        return iou - (c_area - union) / c_area  # GIoU https://arxiv.org/pdf/1902.09630.pdf
    return iou  # IoU

def mask_iou(mask1: torch.Tensor, mask2: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
    """Calculate masks IoU.

    Args:
        mask1 (torch.Tensor): A tensor of shape (N, n) where N is the number of ground truth objects and n is the
            product of image width and height.
        mask2 (torch.Tensor): A tensor of shape (M, n) where M is the number of predicted objects and n is the product
            of image width and height.
        eps (float, optional): A small value to avoid division by zero.

    Returns:
        (torch.Tensor): A tensor of shape (N, M) representing masks IoU.
    """
    intersection = torch.matmul(mask1, mask2.T).clamp_(0)
    union = (mask1.sum(1)[:, None] + mask2.sum(1)[None]) - intersection  # (area1 + area2) - intersection
    return intersection / (union + eps)

def kpt_iou(
    kpt1: torch.Tensor, kpt2: torch.Tensor, area: torch.Tensor, sigma: list[float], eps: float = 1e-7
) -> torch.Tensor:
    """Calculate Object Keypoint Similarity (OKS).

    Args:
        kpt1 (torch.Tensor): A tensor of shape (N, 17, 3) representing ground truth keypoints.
        kpt2 (torch.Tensor): A tensor of shape (M, 17, 3) representing predicted keypoints.
        area (torch.Tensor): A tensor of shape (N,) representing areas from ground truth.
        sigma (list[float]): A list containing 17 values representing keypoint scales.
        eps (float, optional): A small value to avoid division by zero.

    Returns:
        (torch.Tensor): A tensor of shape (N, M) representing keypoint similarities.
    """
    d = (kpt1[:, None, :, 0] - kpt2[..., 0]).pow(2) + (kpt1[:, None, :, 1] - kpt2[..., 1]).pow(2)  # (N, M, 17)
    sigma = torch.tensor(sigma, device=kpt1.device, dtype=kpt1.dtype)  # (17, )
    kpt_mask = kpt1[..., 2] != 0  # (N, 17)
    e = d / ((2 * sigma).pow(2) * (area[:, None, None] + eps) * 2)  # from cocoeval
    # e = d / ((area[None, :, None] + eps) * sigma) ** 2 / 2  # from formula
    return ((-e).exp() * kpt_mask[:, None]).sum(-1) / (kpt_mask.sum(-1)[:, None] + eps)

def _get_covariance_matrix(boxes: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """Generate covariance matrix from oriented bounding boxes.

    Args:
        boxes (torch.Tensor): A tensor of shape (N, 5) representing rotated bounding boxes, with xywhr format.

    Returns:
        (tuple[torch.Tensor, torch.Tensor, torch.Tensor]): Covariance matrix components (a, b, c) where the covariance
            matrix is [[a, c], [c, b]], each of shape (N, 1).
    """
    # Gaussian bounding boxes, ignore the center points (the first two columns) because they are not needed here.
    gbbs = torch.cat((boxes[:, 2:4].pow(2) / 12, boxes[:, 4:]), dim=-1)
    a, b, c = gbbs.split(1, dim=-1)
    cos = c.cos()
    sin = c.sin()
    cos2 = cos.pow(2)
    sin2 = sin.pow(2)
    return a * cos2 + b * sin2, a * sin2 + b * cos2, (a - b) * cos * sin

def probiou(obb1: torch.Tensor, obb2: torch.Tensor, CIoU: bool = False, eps: float = 1e-7) -> torch.Tensor:
    """Calculate probabilistic IoU between oriented bounding boxes.

    Args:
        obb1 (torch.Tensor): Ground truth OBBs, shape (N, 5), format xywhr.
        obb2 (torch.Tensor): Predicted OBBs, shape (N, 5), format xywhr.
        CIoU (bool, optional): If True, calculate CIoU.
        eps (float, optional): Small value to avoid division by zero.

    Returns:
        (torch.Tensor): OBB similarities, shape (N,).

    Notes:
        OBB format: [center_x, center_y, width, height, rotation_angle].

    References:
        https://arxiv.org/pdf/2106.06072v1.pdf
    """
    x1, y1 = obb1[..., :2].split(1, dim=-1)
    x2, y2 = obb2[..., :2].split(1, dim=-1)
    a1, b1, c1 = _get_covariance_matrix(obb1)
    a2, b2, c2 = _get_covariance_matrix(obb2)

    t1 = (
        ((a1 + a2) * (y1 - y2).pow(2) + (b1 + b2) * (x1 - x2).pow(2)) / ((a1 + a2) * (b1 + b2) - (c1 + c2).pow(2) + eps)
    ) * 0.25
    t2 = (((c1 + c2) * (x2 - x1) * (y1 - y2)) / ((a1 + a2) * (b1 + b2) - (c1 + c2).pow(2) + eps)) * 0.5
    t3 = (
        ((a1 + a2) * (b1 + b2) - (c1 + c2).pow(2))
        / (4 * ((a1 * b1 - c1.pow(2)).clamp_(0) * (a2 * b2 - c2.pow(2)).clamp_(0)).sqrt() + eps)
        + eps
    ).log() * 0.5
    bd = (t1 + t2 + t3).clamp(eps, 100.0)
    hd = (1.0 - (-bd).exp() + eps).sqrt()
    iou = 1 - hd
    if CIoU:  # only include the wh aspect ratio part
        w1, h1 = obb1[..., 2:4].split(1, dim=-1)
        w2, h2 = obb2[..., 2:4].split(1, dim=-1)
        v = (4 / math.pi**2) * ((w2 / h2).atan() - (w1 / h1).atan()).pow(2)
        with torch.no_grad():
            alpha = v / (v - iou + (1 + eps))
        return iou - v * alpha  # CIoU
    return iou