COCOMeanAveragePrecision metric

[source]

COCOMeanAveragePrecision class

keras_cv.metrics.COCOMeanAveragePrecision(
    class_ids,
    bounding_box_format,
    recall_thresholds=None,
    iou_thresholds=None,
    area_range=None,
    max_detections=100,
    num_buckets=10000,
    **kwargs
)

COCOMeanAveragePrecision computes an approximation of MaP.

A usage guide is available on keras.io: Using KerasCV COCO metrics. Full implementation details are available in the KerasCV COCO metrics whitepaper.

Arguments

• class_ids: The class IDs to evaluate the metric for. To evaluate for all classes in over a set of sequentially labelled classes, pass range(classes).
• bounding_box_format: Format of the incoming bounding boxes. Supported values are "xywh", "center_xywh", "xyxy".
• iou_thresholds: IoU thresholds over which to evaluate the recall. Must be a tuple of floats, defaults to [0.5:0.05:0.95].
• area_range: area range to constrict the considered bounding boxes in metric computation. Defaults to None, which makes the metric count all bounding boxes. Must be a tuple of floats. The first number in the tuple represents a lower bound for areas, while the second value represents an upper bound. For example, when (0, 32**2) is passed to the metric, recall is only evaluated for objects with areas less than 32*32. If (32**2, 1000000**2) is passed the metric will only be evaluated for boxes with areas larger than 32**2, and smaller than 1000000**2.
• max_detections: number of maximum detections a model is allowed to make. Must be an integer, defaults to 100.
• recall_thresholds: The list of thresholds to average over in the MaP computation. List of floats. Defaults to [0:.01:1].
• num_buckets: num_buckets is used to select the number of confidence buckets predictions are placed into. Instead of computation MaP over each incrementally selected set of bounding boxes, we instead place them into buckets. This makes distributed computation easier. Increasing buckets improves accuracy of the metric, while decreasing buckets improves performance. This is a tradeoff you must weight for your use case. Defaults to 10,000 which is sufficiently large for most use cases.

Usage:

COCOMeanAveragePrecision accepts two Tensors as input to it's update_state() method. These Tensors represent bounding boxes in corners format. Utilities to convert Tensors from xywh to corners format can be found in keras_cv.utils.bounding_box.

Each image in a dataset may have a different number of bounding boxes, both in the ground truth dataset and the prediction set. In order to account for this, you may either pass a tf.RaggedTensor, or pad Tensors with -1s to indicate unused boxes. A utility function to perform this padding is available at keras_cv.bounding_box.pad_batch_to_shape().

coco_map = keras_cv.metrics.COCOMeanAveragePrecision(
    bounding_box_format='xyxy',
    max_detections=100,
    class_ids=[1]
)

y_true = np.array([[[0, 0, 10, 10, 1], [20, 20, 10, 10, 1]]]).astype(np.float32)
y_pred = np.array([[[0, 0, 10, 10, 1, 1.0], [5, 5, 10, 10, 1, 0.9]]]).astype(
    np.float32
)
coco_map.update_state(y_true, y_pred)
coco_map.result()
# 0.24752477