Skip to content

Detection Metrics

Warning

Evaluation API is still fluid and may change. If you use Evaluation API in your project until further notice, freeze the supervision version in your requirements.txt or setup.py.

Confusion matrix for object detection tasks.

Attributes:

Name Type Description
matrix ndarray

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.

Source code in supervision/metrics/detection.py
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
@dataclass
class ConfusionMatrix:
    """
    Confusion matrix for object detection tasks.

    Attributes:
        matrix (np.ndarray): 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`.
    """

    matrix: np.ndarray
    classes: List[str]
    conf_threshold: float
    iou_threshold: float

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

        Args:
            targets (List[Detections]): Detections objects from ground-truth.
            predictions (List[Detections]): Detections objects predicted by the model.
            classes (List[str]): Model class names.
            conf_threshold (float): Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold (float): Detection IoU threshold between `0` and `1`.
                Detections with lower IoU will be classified as `FP`.

        Returns:
            ConfusionMatrix: New instance of ConfusionMatrix.

        Example:
            ```python
            import supervision as sv

            targets = [
                sv.Detections(...),
                sv.Detections(...)
            ]

            predictions = [
                sv.Detections(...),
                sv.Detections(...)
            ]

            confusion_matrix = sv.ConfusionMatrix.from_detections(
                predictions=predictions,
                targets=target,
                classes=['person', ...]
            )

            print(confusion_matrix.matrix)
            # np.array([
            #    [0., 0., 0., 0.],
            #    [0., 1., 0., 1.],
            #    [0., 1., 1., 0.],
            #    [1., 1., 0., 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,
            classes=classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
        )

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

        Args:
            predictions (List[np.ndarray]): 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 (List[np.ndarray]): 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.
            classes (List[str]): Model class names.
            conf_threshold (float): Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold (float): Detection iou  threshold between `0` and `1`.
                Detections with lower iou will be classified as `FP`.

        Returns:
            ConfusionMatrix: New instance of ConfusionMatrix.

        Example:
            ```python
            import supervision as sv
            import numpy as np

            targets = (
                [
                    np.array(
                        [
                            [0.0, 0.0, 3.0, 3.0, 1],
                            [2.0, 2.0, 5.0, 5.0, 1],
                            [6.0, 1.0, 8.0, 3.0, 2],
                        ]
                    ),
                    np.array([1.0, 1.0, 2.0, 2.0, 2]),
                ]
            )

            predictions = [
                np.array(
                    [
                        [0.0, 0.0, 3.0, 3.0, 1, 0.9],
                        [0.1, 0.1, 3.0, 3.0, 0, 0.9],
                        [6.0, 1.0, 8.0, 3.0, 1, 0.8],
                        [1.0, 6.0, 2.0, 7.0, 1, 0.8],
                    ]
                ),
                np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
            ]

            confusion_matrix = sv.ConfusionMatrix.from_tensors(
                predictions=predictions,
                targets=targets,
                classes=['person', ...]
            )

            print(confusion_matrix.matrix)
            # np.array([
            #     [0., 0., 0., 0.],
            #     [0., 1., 0., 1.],
            #     [0., 1., 1., 0.],
            #     [1., 1., 0., 0.]
            # ])
            ```
        """
        validate_input_tensors(predictions, targets)

        num_classes = len(classes)
        matrix = np.zeros((num_classes + 1, num_classes + 1))
        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,
            )
        return cls(
            matrix=matrix,
            classes=classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
        )

    @staticmethod
    def evaluate_detection_batch(
        predictions: np.ndarray,
        targets: np.ndarray,
        num_classes: int,
        conf_threshold: float,
        iou_threshold: float,
    ) -> np.ndarray:
        """
        Calculate confusion matrix for a batch of detections for a single image.

        Args:
            predictions (np.ndarray): 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 (np.ndarray): 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.
            num_classes (int): Number of classes.
            conf_threshold (float): Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold (float): Detection iou  threshold between `0` and `1`.
                Detections with lower iou will be classified as `FP`.

        Returns:
            np.ndarray: Confusion matrix based on a single image.
        """
        result_matrix = np.zeros((num_classes + 1, num_classes + 1))

        conf_idx = 5
        confidence = predictions[:, conf_idx]
        detection_batch_filtered = predictions[confidence > conf_threshold]

        class_id_idx = 4
        true_classes = np.array(targets[:, class_id_idx], dtype=np.int16)
        detection_classes = np.array(
            detection_batch_filtered[:, class_id_idx], dtype=np.int16
        )
        true_boxes = targets[:, :class_id_idx]
        detection_boxes = detection_batch_filtered[:, :class_id_idx]

        iou_batch = box_iou_batch(
            boxes_true=true_boxes, boxes_detection=detection_boxes
        )
        matched_idx = np.asarray(iou_batch > iou_threshold).nonzero()

        if matched_idx[0].shape[0]:
            matches = np.stack(
                (matched_idx[0], matched_idx[1], iou_batch[matched_idx]), axis=1
            )
            matches = ConfusionMatrix._drop_extra_matches(matches=matches)
        else:
            matches = np.zeros((0, 3))

        matched_true_idx, matched_detection_idx, _ = matches.transpose().astype(
            np.int16
        )

        for i, true_class_value in enumerate(true_classes):
            j = matched_true_idx == i
            if matches.shape[0] > 0 and sum(j) == 1:
                result_matrix[
                    true_class_value, detection_classes[matched_detection_idx[j]]
                ] += 1  # TP
            else:
                result_matrix[true_class_value, num_classes] += 1  # FN

        for i, detection_class_value in enumerate(detection_classes):
            if not any(matched_detection_idx == i):
                result_matrix[num_classes, detection_class_value] += 1  # FP

        return result_matrix

    @staticmethod
    def _drop_extra_matches(matches: np.ndarray) -> np.ndarray:
        """
        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]]
        return matches

    @classmethod
    def benchmark(
        cls,
        dataset: DetectionDataset,
        callback: Callable[[np.ndarray], Detections],
        conf_threshold: float = 0.3,
        iou_threshold: float = 0.5,
    ) -> ConfusionMatrix:
        """
        Calculate confusion matrix from dataset and callback function.

        Args:
            dataset (DetectionDataset): Object detection dataset used for evaluation.
            callback (Callable[[np.ndarray], Detections]): Function that takes an image
                as input and returns Detections object.
            conf_threshold (float): Detection confidence threshold between `0` and `1`.
                Detections with lower confidence will be excluded.
            iou_threshold (float): Detection IoU threshold between `0` and `1`.
                Detections with lower IoU will be classified as `FP`.

        Returns:
            ConfusionMatrix: 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.]
            # ])
            ```
        """
        predictions, targets = [], []
        for img_name, img in dataset.images.items():
            predictions_batch = callback(img)
            predictions.append(predictions_batch)
            targets_batch = dataset.annotations[img_name]
            targets.append(targets_batch)
        return cls.from_detections(
            predictions=predictions,
            targets=targets,
            classes=dataset.classes,
            conf_threshold=conf_threshold,
            iou_threshold=iou_threshold,
        )

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

        Args:
            save_path (Optional[str]): Path to save the plot. If not provided,
                plot will be displayed.
            title (Optional[str]): Title of the plot.
            classes (Optional[List[str]]): List of classes to be displayed on the plot.
                If not provided, all classes will be displayed.
            normalize (bool): If True, normalize the confusion matrix.
            fig_size (Tuple[int, int]): Size of the plot.

        Returns:
            matplotlib.figure.Figure: Confusion matrix plot.
        """

        array = self.matrix.copy()

        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=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

Functions

benchmark(dataset, callback, conf_threshold=0.3, iou_threshold=0.5) 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], Detections]

Function that takes an image as input and returns 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

Returns:

Name Type Description
ConfusionMatrix 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 supervision/metrics/detection.py
@classmethod
def benchmark(
    cls,
    dataset: DetectionDataset,
    callback: Callable[[np.ndarray], Detections],
    conf_threshold: float = 0.3,
    iou_threshold: float = 0.5,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix from dataset and callback function.

    Args:
        dataset (DetectionDataset): Object detection dataset used for evaluation.
        callback (Callable[[np.ndarray], Detections]): Function that takes an image
            as input and returns Detections object.
        conf_threshold (float): Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold (float): Detection IoU threshold between `0` and `1`.
            Detections with lower IoU will be classified as `FP`.

    Returns:
        ConfusionMatrix: 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.]
        # ])
        ```
    """
    predictions, targets = [], []
    for img_name, img in dataset.images.items():
        predictions_batch = callback(img)
        predictions.append(predictions_batch)
        targets_batch = dataset.annotations[img_name]
        targets.append(targets_batch)
    return cls.from_detections(
        predictions=predictions,
        targets=targets,
        classes=dataset.classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
    )

evaluate_detection_batch(predictions, targets, num_classes, conf_threshold, iou_threshold) staticmethod

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

Parameters:

Name Type Description Default
predictions ndarray

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.

required
targets ndarray

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.

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

Returns:

Type Description
ndarray

np.ndarray: Confusion matrix based on a single image.

Source code in supervision/metrics/detection.py
@staticmethod
def evaluate_detection_batch(
    predictions: np.ndarray,
    targets: np.ndarray,
    num_classes: int,
    conf_threshold: float,
    iou_threshold: float,
) -> np.ndarray:
    """
    Calculate confusion matrix for a batch of detections for a single image.

    Args:
        predictions (np.ndarray): 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 (np.ndarray): 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.
        num_classes (int): Number of classes.
        conf_threshold (float): Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold (float): Detection iou  threshold between `0` and `1`.
            Detections with lower iou will be classified as `FP`.

    Returns:
        np.ndarray: Confusion matrix based on a single image.
    """
    result_matrix = np.zeros((num_classes + 1, num_classes + 1))

    conf_idx = 5
    confidence = predictions[:, conf_idx]
    detection_batch_filtered = predictions[confidence > conf_threshold]

    class_id_idx = 4
    true_classes = np.array(targets[:, class_id_idx], dtype=np.int16)
    detection_classes = np.array(
        detection_batch_filtered[:, class_id_idx], dtype=np.int16
    )
    true_boxes = targets[:, :class_id_idx]
    detection_boxes = detection_batch_filtered[:, :class_id_idx]

    iou_batch = box_iou_batch(
        boxes_true=true_boxes, boxes_detection=detection_boxes
    )
    matched_idx = np.asarray(iou_batch > iou_threshold).nonzero()

    if matched_idx[0].shape[0]:
        matches = np.stack(
            (matched_idx[0], matched_idx[1], iou_batch[matched_idx]), axis=1
        )
        matches = ConfusionMatrix._drop_extra_matches(matches=matches)
    else:
        matches = np.zeros((0, 3))

    matched_true_idx, matched_detection_idx, _ = matches.transpose().astype(
        np.int16
    )

    for i, true_class_value in enumerate(true_classes):
        j = matched_true_idx == i
        if matches.shape[0] > 0 and sum(j) == 1:
            result_matrix[
                true_class_value, detection_classes[matched_detection_idx[j]]
            ] += 1  # TP
        else:
            result_matrix[true_class_value, num_classes] += 1  # FN

    for i, detection_class_value in enumerate(detection_classes):
        if not any(matched_detection_idx == i):
            result_matrix[num_classes, detection_class_value] += 1  # FP

    return result_matrix

from_detections(predictions, targets, classes, conf_threshold=0.3, iou_threshold=0.5) 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

Returns:

Name Type Description
ConfusionMatrix ConfusionMatrix

New instance of ConfusionMatrix.

Example
import supervision as sv

targets = [
    sv.Detections(...),
    sv.Detections(...)
]

predictions = [
    sv.Detections(...),
    sv.Detections(...)
]

confusion_matrix = sv.ConfusionMatrix.from_detections(
    predictions=predictions,
    targets=target,
    classes=['person', ...]
)

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 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,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix based on predicted and ground-truth detections.

    Args:
        targets (List[Detections]): Detections objects from ground-truth.
        predictions (List[Detections]): Detections objects predicted by the model.
        classes (List[str]): Model class names.
        conf_threshold (float): Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold (float): Detection IoU threshold between `0` and `1`.
            Detections with lower IoU will be classified as `FP`.

    Returns:
        ConfusionMatrix: New instance of ConfusionMatrix.

    Example:
        ```python
        import supervision as sv

        targets = [
            sv.Detections(...),
            sv.Detections(...)
        ]

        predictions = [
            sv.Detections(...),
            sv.Detections(...)
        ]

        confusion_matrix = sv.ConfusionMatrix.from_detections(
            predictions=predictions,
            targets=target,
            classes=['person', ...]
        )

        print(confusion_matrix.matrix)
        # np.array([
        #    [0., 0., 0., 0.],
        #    [0., 1., 0., 1.],
        #    [0., 1., 1., 0.],
        #    [1., 1., 0., 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,
        classes=classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
    )

from_tensors(predictions, targets, classes, conf_threshold=0.3, iou_threshold=0.5) classmethod

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

Parameters:

Name Type Description Default
predictions List[ndarray]

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]

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.

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

Returns:

Name Type Description
ConfusionMatrix ConfusionMatrix

New instance of ConfusionMatrix.

Example
import supervision as sv
import numpy as np

targets = (
    [
        np.array(
            [
                [0.0, 0.0, 3.0, 3.0, 1],
                [2.0, 2.0, 5.0, 5.0, 1],
                [6.0, 1.0, 8.0, 3.0, 2],
            ]
        ),
        np.array([1.0, 1.0, 2.0, 2.0, 2]),
    ]
)

predictions = [
    np.array(
        [
            [0.0, 0.0, 3.0, 3.0, 1, 0.9],
            [0.1, 0.1, 3.0, 3.0, 0, 0.9],
            [6.0, 1.0, 8.0, 3.0, 1, 0.8],
            [1.0, 6.0, 2.0, 7.0, 1, 0.8],
        ]
    ),
    np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
]

confusion_matrix = sv.ConfusionMatrix.from_tensors(
    predictions=predictions,
    targets=targets,
    classes=['person', ...]
)

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 supervision/metrics/detection.py
@classmethod
def from_tensors(
    cls,
    predictions: List[np.ndarray],
    targets: List[np.ndarray],
    classes: List[str],
    conf_threshold: float = 0.3,
    iou_threshold: float = 0.5,
) -> ConfusionMatrix:
    """
    Calculate confusion matrix based on predicted and ground-truth detections.

    Args:
        predictions (List[np.ndarray]): 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 (List[np.ndarray]): 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.
        classes (List[str]): Model class names.
        conf_threshold (float): Detection confidence threshold between `0` and `1`.
            Detections with lower confidence will be excluded.
        iou_threshold (float): Detection iou  threshold between `0` and `1`.
            Detections with lower iou will be classified as `FP`.

    Returns:
        ConfusionMatrix: New instance of ConfusionMatrix.

    Example:
        ```python
        import supervision as sv
        import numpy as np

        targets = (
            [
                np.array(
                    [
                        [0.0, 0.0, 3.0, 3.0, 1],
                        [2.0, 2.0, 5.0, 5.0, 1],
                        [6.0, 1.0, 8.0, 3.0, 2],
                    ]
                ),
                np.array([1.0, 1.0, 2.0, 2.0, 2]),
            ]
        )

        predictions = [
            np.array(
                [
                    [0.0, 0.0, 3.0, 3.0, 1, 0.9],
                    [0.1, 0.1, 3.0, 3.0, 0, 0.9],
                    [6.0, 1.0, 8.0, 3.0, 1, 0.8],
                    [1.0, 6.0, 2.0, 7.0, 1, 0.8],
                ]
            ),
            np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
        ]

        confusion_matrix = sv.ConfusionMatrix.from_tensors(
            predictions=predictions,
            targets=targets,
            classes=['person', ...]
        )

        print(confusion_matrix.matrix)
        # np.array([
        #     [0., 0., 0., 0.],
        #     [0., 1., 0., 1.],
        #     [0., 1., 1., 0.],
        #     [1., 1., 0., 0.]
        # ])
        ```
    """
    validate_input_tensors(predictions, targets)

    num_classes = len(classes)
    matrix = np.zeros((num_classes + 1, num_classes + 1))
    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,
        )
    return cls(
        matrix=matrix,
        classes=classes,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
    )

plot(save_path=None, title=None, classes=None, normalize=False, fig_size=(12, 10))

Create confusion matrix plot and save it at selected location.

Parameters:

Name Type Description Default
save_path Optional[str]

Path to save the plot. If not provided, plot will be displayed.

None
title Optional[str]

Title of the plot.

None
classes Optional[List[str]]

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

matplotlib.figure.Figure: Confusion matrix plot.

Source code in supervision/metrics/detection.py
def plot(
    self,
    save_path: Optional[str] = None,
    title: Optional[str] = None,
    classes: Optional[List[str]] = None,
    normalize: bool = False,
    fig_size: Tuple[int, int] = (12, 10),
) -> matplotlib.figure.Figure:
    """
    Create confusion matrix plot and save it at selected location.

    Args:
        save_path (Optional[str]): Path to save the plot. If not provided,
            plot will be displayed.
        title (Optional[str]): Title of the plot.
        classes (Optional[List[str]]): List of classes to be displayed on the plot.
            If not provided, all classes will be displayed.
        normalize (bool): If True, normalize the confusion matrix.
        fig_size (Tuple[int, int]): Size of the plot.

    Returns:
        matplotlib.figure.Figure: Confusion matrix plot.
    """

    array = self.matrix.copy()

    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=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

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

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 supervision/metrics/detection.py
@dataclass(frozen=True)
class MeanAveragePrecision:
    """
    Mean Average Precision for object detection tasks.

    Attributes:
        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 (np.ndarray): 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: np.ndarray

    @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 (List[Detections]): Detections objects from ground-truth.
            predictions (List[Detections]): Detections objects predicted by the model.
        Returns:
            MeanAveragePrecision: New instance of ConfusionMatrix.

        Example:
            ```python
            import supervision as sv

            targets = [
                sv.Detections(...),
                sv.Detections(...)
            ]

            predictions = [
                sv.Detections(...),
                sv.Detections(...)
            ]

            mean_average_precision = sv.MeanAveragePrecision.from_detections(
                predictions=predictions,
                targets=target,
            )

            print(mean_average_precison.map50_95)
            # 0.2899
            ```
        """
        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[[np.ndarray], Detections],
    ) -> MeanAveragePrecision:
        """
        Calculate mean average precision from dataset and callback function.

        Args:
            dataset (DetectionDataset): Object detection dataset used for evaluation.
            callback (Callable[[np.ndarray], Detections]): Function that takes
                an image as input and returns Detections object.
        Returns:
            MeanAveragePrecision: 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 img_name, img in dataset.images.items():
            predictions_batch = callback(img)
            predictions.append(predictions_batch)
            targets_batch = dataset.annotations[img_name]
            targets.append(targets_batch)
        return cls.from_detections(
            predictions=predictions,
            targets=targets,
        )

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

        Args:
            predictions (List[np.ndarray]): 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 (List[np.ndarray]): 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.
        Returns:
            MeanAveragePrecision: New instance of MeanAveragePrecision.

        Example:
            ```python
            import supervision as sv
            import numpy as np

            targets = (
                [
                    np.array(
                        [
                            [0.0, 0.0, 3.0, 3.0, 1],
                            [2.0, 2.0, 5.0, 5.0, 1],
                            [6.0, 1.0, 8.0, 3.0, 2],
                        ]
                    ),
                    np.array([[1.0, 1.0, 2.0, 2.0, 2]]),
                ]
            )

            predictions = [
                np.array(
                    [
                        [0.0, 0.0, 3.0, 3.0, 1, 0.9],
                        [0.1, 0.1, 3.0, 3.0, 0, 0.9],
                        [6.0, 1.0, 8.0, 3.0, 1, 0.8],
                        [1.0, 6.0, 2.0, 7.0, 1, 0.8],
                    ]
                ),
                np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
            ]

            mean_average_precison = sv.MeanAveragePrecision.from_tensors(
                predictions=predictions,
                targets=targets,
            )

            print(mean_average_precison.map50_95)
            # 0.6649
            ```
        """
        validate_input_tensors(predictions, targets)
        iou_thresholds = np.linspace(0.5, 0.95, 10)
        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],
                    )
                )

        # 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(*concatenated_stats)
            map50 = average_precisions[:, 0].mean()
            map75 = average_precisions[:, 5].mean()
            map50_95 = average_precisions.mean()
        else:
            map50, map75, map50_95 = 0, 0, 0
            average_precisions = []

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

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

        Args:
            recall (np.ndarray): The recall curve.
            precision (np.ndarray): The precision curve.

        Returns:
            float: Average precision.
        """
        extended_recall = np.concatenate(([0.0], recall, [1.0]))
        extended_precision = np.concatenate(([1.0], precision, [0.0]))
        max_accumulated_precision = np.flip(
            np.maximum.accumulate(np.flip(extended_precision))
        )
        interpolated_recall_levels = np.linspace(0, 1, 101)
        interpolated_precision = np.interp(
            interpolated_recall_levels, extended_recall, max_accumulated_precision
        )
        average_precision = np.trapz(interpolated_precision, interpolated_recall_levels)
        return average_precision

    @staticmethod
    def _match_detection_batch(
        predictions: np.ndarray, targets: np.ndarray, iou_thresholds: np.ndarray
    ) -> np.ndarray:
        """
        Match predictions with target labels based on IoU levels.

        Args:
            predictions (np.ndarray): 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 (np.ndarray): 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 (np.ndarray): Array contains different IoU thresholds.

        Returns:
            np.ndarray: 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)

            if matched_indices[0].shape[0]:
                combined_indices = np.stack(matched_indices, axis=1)
                iou_values = iou[matched_indices][:, None]
                matches = np.hstack([combined_indices, iou_values])

                if matched_indices[0].shape[0] > 1:
                    matches = matches[matches[:, 2].argsort()[::-1]]
                    matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
                    matches = matches[np.unique(matches[:, 0], return_index=True)[1]]

                correct[matches[:, 1].astype(int), i] = True

        return correct

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

        Args:
            matches (np.ndarray): True positives.
            prediction_confidence (np.ndarray): Objectness value from 0-1.
            prediction_class_ids (np.ndarray): Predicted object classes.
            true_class_ids (np.ndarray): True object classes.
            eps (float, optional): Small value to prevent division by zero.

        Returns:
            np.ndarray: 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 = np.zeros((num_classes, matches.shape[1]))

        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]
                    )
                )

        return average_precisions

Functions

benchmark(dataset, callback) 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], Detections]

Function that takes an image as input and returns Detections object.

required

Returns: MeanAveragePrecision: 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 supervision/metrics/detection.py
@classmethod
def benchmark(
    cls,
    dataset: DetectionDataset,
    callback: Callable[[np.ndarray], Detections],
) -> MeanAveragePrecision:
    """
    Calculate mean average precision from dataset and callback function.

    Args:
        dataset (DetectionDataset): Object detection dataset used for evaluation.
        callback (Callable[[np.ndarray], Detections]): Function that takes
            an image as input and returns Detections object.
    Returns:
        MeanAveragePrecision: 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 img_name, img in dataset.images.items():
        predictions_batch = callback(img)
        predictions.append(predictions_batch)
        targets_batch = dataset.annotations[img_name]
        targets.append(targets_batch)
    return cls.from_detections(
        predictions=predictions,
        targets=targets,
    )

compute_average_precision(recall, precision) staticmethod

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

Parameters:

Name Type Description Default
recall ndarray

The recall curve.

required
precision ndarray

The precision curve.

required

Returns:

Name Type Description
float float

Average precision.

Source code in supervision/metrics/detection.py
@staticmethod
def compute_average_precision(recall: np.ndarray, precision: np.ndarray) -> float:
    """
    Compute the average precision using 101-point interpolation (COCO), given
        the recall and precision curves.

    Args:
        recall (np.ndarray): The recall curve.
        precision (np.ndarray): The precision curve.

    Returns:
        float: Average precision.
    """
    extended_recall = np.concatenate(([0.0], recall, [1.0]))
    extended_precision = np.concatenate(([1.0], precision, [0.0]))
    max_accumulated_precision = np.flip(
        np.maximum.accumulate(np.flip(extended_precision))
    )
    interpolated_recall_levels = np.linspace(0, 1, 101)
    interpolated_precision = np.interp(
        interpolated_recall_levels, extended_recall, max_accumulated_precision
    )
    average_precision = np.trapz(interpolated_precision, interpolated_recall_levels)
    return average_precision

from_detections(predictions, targets) 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: MeanAveragePrecision: New instance of ConfusionMatrix.

Example
import supervision as sv

targets = [
    sv.Detections(...),
    sv.Detections(...)
]

predictions = [
    sv.Detections(...),
    sv.Detections(...)
]

mean_average_precision = sv.MeanAveragePrecision.from_detections(
    predictions=predictions,
    targets=target,
)

print(mean_average_precison.map50_95)
# 0.2899
Source code in 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 (List[Detections]): Detections objects from ground-truth.
        predictions (List[Detections]): Detections objects predicted by the model.
    Returns:
        MeanAveragePrecision: New instance of ConfusionMatrix.

    Example:
        ```python
        import supervision as sv

        targets = [
            sv.Detections(...),
            sv.Detections(...)
        ]

        predictions = [
            sv.Detections(...),
            sv.Detections(...)
        ]

        mean_average_precision = sv.MeanAveragePrecision.from_detections(
            predictions=predictions,
            targets=target,
        )

        print(mean_average_precison.map50_95)
        # 0.2899
        ```
    """
    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, targets) classmethod

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

Parameters:

Name Type Description Default
predictions List[ndarray]

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]

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.

required

Returns: MeanAveragePrecision: New instance of MeanAveragePrecision.

Example
import supervision as sv
import numpy as np

targets = (
    [
        np.array(
            [
                [0.0, 0.0, 3.0, 3.0, 1],
                [2.0, 2.0, 5.0, 5.0, 1],
                [6.0, 1.0, 8.0, 3.0, 2],
            ]
        ),
        np.array([[1.0, 1.0, 2.0, 2.0, 2]]),
    ]
)

predictions = [
    np.array(
        [
            [0.0, 0.0, 3.0, 3.0, 1, 0.9],
            [0.1, 0.1, 3.0, 3.0, 0, 0.9],
            [6.0, 1.0, 8.0, 3.0, 1, 0.8],
            [1.0, 6.0, 2.0, 7.0, 1, 0.8],
        ]
    ),
    np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
]

mean_average_precison = sv.MeanAveragePrecision.from_tensors(
    predictions=predictions,
    targets=targets,
)

print(mean_average_precison.map50_95)
# 0.6649
Source code in supervision/metrics/detection.py
@classmethod
def from_tensors(
    cls,
    predictions: List[np.ndarray],
    targets: List[np.ndarray],
) -> MeanAveragePrecision:
    """
    Calculate Mean Average Precision based on predicted and ground-truth
        detections at different threshold.

    Args:
        predictions (List[np.ndarray]): 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 (List[np.ndarray]): 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.
    Returns:
        MeanAveragePrecision: New instance of MeanAveragePrecision.

    Example:
        ```python
        import supervision as sv
        import numpy as np

        targets = (
            [
                np.array(
                    [
                        [0.0, 0.0, 3.0, 3.0, 1],
                        [2.0, 2.0, 5.0, 5.0, 1],
                        [6.0, 1.0, 8.0, 3.0, 2],
                    ]
                ),
                np.array([[1.0, 1.0, 2.0, 2.0, 2]]),
            ]
        )

        predictions = [
            np.array(
                [
                    [0.0, 0.0, 3.0, 3.0, 1, 0.9],
                    [0.1, 0.1, 3.0, 3.0, 0, 0.9],
                    [6.0, 1.0, 8.0, 3.0, 1, 0.8],
                    [1.0, 6.0, 2.0, 7.0, 1, 0.8],
                ]
            ),
            np.array([[1.0, 1.0, 2.0, 2.0, 2, 0.8]])
        ]

        mean_average_precison = sv.MeanAveragePrecision.from_tensors(
            predictions=predictions,
            targets=targets,
        )

        print(mean_average_precison.map50_95)
        # 0.6649
        ```
    """
    validate_input_tensors(predictions, targets)
    iou_thresholds = np.linspace(0.5, 0.95, 10)
    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],
                )
            )

    # 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(*concatenated_stats)
        map50 = average_precisions[:, 0].mean()
        map75 = average_precisions[:, 5].mean()
        map50_95 = average_precisions.mean()
    else:
        map50, map75, map50_95 = 0, 0, 0
        average_precisions = []

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

Comments