Table of Contents
Contents

fiftyone.utils.iou

Intersection over union (IoU) utilities.

Copyright 2017-2023, Voxel51, Inc.
voxel51.com

Functions:

compute_ious(preds, gts[, iscrowd, …])

Computes the pairwise IoUs between the predicted and ground truth objects.

compute_max_ious(sample_collection, label_field)

Populates an attribute on each label in the given spatial field(s) that records the max IoU between the object and another object in the same sample/frame.

compute_segment_ious(preds, gts)

Computes the pairwise IoUs between the predicted and ground truth temporal detections.

find_duplicates(sample_collection, label_field)

Returns IDs of duplicate labels in the given field of the collection, as defined as labels with an IoU greater than a chosen threshold with another label in the field.
fiftyone.utils.iou.compute_ious(preds, gts, iscrowd=None, classwise=False, use_masks=False, use_boxes=False, tolerance=None, error_level=1)

Computes the pairwise IoUs between the predicted and ground truth objects.

For polylines, IoUs are computed assuming the shapes are solid (filled), regardless of their filled attributes.

For keypoints, “IoUs” are computed via object keypoint similarity.

Parameters

  • preds – a list of predicted fiftyone.core.labels.Detection, fiftyone.core.labels.Polyline, or fiftyone.core.labels.Keypoints instances

  • gts – a list of ground truth fiftyone.core.labels.Detection, fiftyone.core.labels.Polyline, or fiftyone.core.labels.Keypoints instances

  • iscrowd (None) – an optional name of a boolean attribute or boolean function to apply to each label that determines whether a ground truth object is a crowd. If provided, the area of the predicted object is used as the “union” area for IoU calculations involving crowd objects

  • classwise (False) – whether to consider objects with different label values as always non-overlapping (True) or to compute IoUs for all objects regardless of label (False)

  • use_masks (False) – whether to compute IoUs using the instances masks in the mask attribute of the provided objects, which must be fiftyone.core.labels.Detection instances

  • use_boxes (False) – whether to compute IoUs using the bounding boxes of the provided fiftyone.core.labels.Polyline instances rather than using their actual geometries

  • tolerance (None) – a tolerance, in pixels, when generating approximate polylines for instance masks. Typical values are 1-3 pixels

  • error_level (1) –

    the error level to use when manipulating instance masks or polylines. Valid values are:

    • 0: raise geometric errors that are encountered

    • 1: log warnings if geometric errors are encountered

    • 2: ignore geometric errors

    If error_level > 0, any calculation that raises a geometric error will default to an IoU of 0

Returns

a num_preds x num_gts array of IoUs

fiftyone.utils.iou.compute_segment_ious(preds, gts)

Computes the pairwise IoUs between the predicted and ground truth temporal detections.

Parameters
Returns

a num_preds x num_gts array of segment IoUs

fiftyone.utils.iou.compute_max_ious(sample_collection, label_field, other_field=None, iou_attr='max_iou', id_attr=None, **kwargs)

Populates an attribute on each label in the given spatial field(s) that records the max IoU between the object and another object in the same sample/frame.

If other_field is provided, IoUs are computed between objects in label_field and other_field.

If no other_field is provided, IoUs are computed between objects in label_field alone, excluding the object itself.

Parameters
fiftyone.utils.iou.find_duplicates(sample_collection, label_field, iou_thresh=0.999, method='simple', **kwargs)

Returns IDs of duplicate labels in the given field of the collection, as defined as labels with an IoU greater than a chosen threshold with another label in the field.

The following duplicate removal methods are supported:

  • method="simple": mark the latter label in every pair of labels whose IoU exceeds the threshold. This may remove more labels than the greedy method

  • method="greedy": apply a greedy method to mark the fewest number of labels as duplicate such that no non-duplicate labels have IoU greater than the specified threshold. This method is more computationally expensive than the simple method

Parameters
Returns

a list of IDs of duplicate labels