Documentation for detection_utils.metrics

compute_precision(prediction_detections, …)	Compute the average precision of predictions given targets.
compute_recall(prediction_detections, …)	Compute the average recall of predictions given targets.

detection_utils.metrics.compute_precision(prediction_detections: numpy.ndarray, truth_detections: numpy.ndarray, threshold: float = 0.5) → float

Compute the average precision of predictions given targets.

Precision is defined as the number of true positive predictions divided by the number of total positive predictions.

Parameters:

prediction_detections : numpy.ndarray, shape=(N, 5)

The predicted objects, in (left, top, right, bottom, class) format.

truth_detections : numpy.ndarray, shape=(K, 5)

The ground-truth objects in (left, top, right, bottom, class) format.

threshold : Real, optional (default=0.5)

The IoU threshold at which to compute precision.

Returns:

float

The average precision (AP) for the given detections and truth.

Notes

This function operates such that when there are zero predictions, precision is 1.

Examples

>>> from detection_utils.metrics import compute_precision
>>> import numpy as np
>>> predictions = np.array([[0, 0, 10, 10, 1], [3, 3, 7, 7, 1]])  # left, top, right, bottom, class prediction
>>> actual = np.array([[2, 3, 6, 7, 1]])
>>> compute_precision(predictions, actual)
0.5

# Our IoUs are 0.16, 0.6 >>> compute_precision(predictions, actual, threshold=0.15) 1.0

>>> compute_precision(predictions, actual, threshold=0.75)
0.0

detection_utils.metrics.compute_recall(prediction_detections: numpy.ndarray, truth_detections: numpy.ndarray, threshold: float = 0.5) → float

Compute the average recall of predictions given targets.

Recall is defined as the number true positive predictions divided by the number of ground-truth targets.

Parameters:

prediction_detections : numpy.ndarray, shape=(N, 5)

The predicted objects, in (left, top, right, bottom, class) format.

truth_detections : numpy.ndarray, shape=(K, 5)

The ground-truth objects in (left, top, right, bottom, class) format.

threshold : Real, optional (default=0.5)

The IoU threshold at which to compute recall.

Returns:

float

The average recall (AR) for the given detections and truth.

Notes

This function operates such that when there are zero targets, recall is 1 regardless of predictions.

Examples

>>> from detection_utils.metrics import compute_recall
>>> import numpy as np
>>> predictions = np.array([[0, 0, 10, 10, 1], [3, 3, 7, 7, 1]])  # left, top, right, bottom, class prediction
>>> actual = np.array([[2, 3, 6, 7, 1]])
>>> compute_recall(predictions, actual)
1.0

# Our highest IoU is 0.6 so let’s set our threshold above that >>> compute_recall(predictions, actual, threshold=0.75) 0.0