Table 4 An overview of the 12 COCO metrics.

From: GooseDetectlion: A Fully Annotated Dataset for Lion-head Goose Detection in Smart Farms

Metric

Specification

AP (Average Precision)

mAP

The average of the AP over 10 IoU thresholds of .50:.05:.95, where AP is the mean aver-age precision over all categories, and IoU is the intersection over union

AP50

The AP at the IoU thresholds of .50, this metric is equivalent to the Pascal VOC metric

AP75

The AP at the IoU thresholds of .75

AP Across Scales

APS

The AP for small objects with bounding box area less than 322 pixels

APM

The AP for medium objects with bounding box area between 322 and 962 pixels

APL

The AP for large objects with bounding box area greater than 962 pixels

AR (Average Recall)

AR1

The maximum recall given 1 detection per image, which is averaged over all categories and IoUs

AR10

The maximum recall given 10 detections per image, which is averaged over all categories and IoUs

AR100

The maximum recall given 100 detections per image, which is averaged over all categories and IoUs

AR Across Scales

ARS

The AR for small objects with bounding box area less than 322 pixels

ARM

The AR for medium objects with bounding box area between 322 and 962 pixels

ARL

The AR for large objects with bounding box area greater than 962 pixels