pairwise_pixel_stats

Compute the # of TP, FP, FN pixels for each object in a labelled/semantic mask.

Optionally a binary metric can be computed instead of the satistics. Atleast 2x faster than computing with np.histogram2d.

Parameters

true : np.ndarray
    Ground truth (semantic or labelled mask). Shape (H, W).
pred : np.ndarray
    Predicted (semantic or labelled mask). Shape (H, W).
num_classes : int, optional
    Number of classes in the dataset. If None, stats are computed for instances.
    If not None stats are computed for classes i.e. semantic segmentation masks.
metric_func : Callable, optional
    A binary metric function. e.g. `iou_score` or `dice`.

Returns

List[np.ndarray, ...] or None:
    A List of 2D arrays (i, j) where i corresponds to a ground
    truth label and j corresponds to a predicted label. Each value
    of the matrix is the computed statistic or metric at pos (i, j).
    By default. returns the tp, fp, and fn matrices.

    If stats computed for instances:
        Shape: (n_labels_gt, n_labels_pred). Dtype. float64.
    If stats computed for classes:
        Shape: (num_classes, num_classes). Dtype. float64.