*3.3. Evaluation Criteria*

The COCO metrics [67] are adopted. Its core index is the average precision (*AP*), i.e., the average value of precisions under ten intersection over union (IOU) thresholds from 0.50 to 0.95 with an interval of 0.05. *AP*50 denotes the *AP* of an *IOU* threshold of 0.50. *AP*75 denotes *AP* of an *IOU* threshold of 0.75. *APS* denotes *AP* of small ships (<322 pixels). *APM* denotes the *AP* of medium ships (>322 pixels and <962 pixels). *APL* denotes the *AP* of large ships (>962 pixels). Specifically, the *IOU* of the predicted mask and the ground truth mask is described by

$$IOLI = \frac{P\_{mask} \cap G\_{mask}}{P\_{mask} \cup G\_{mask}} \tag{10}$$

where *Pmask* represents the predicted mask and *Gmask* represents the ground truth mask. According to a given *IOU* threshold and confidence threshold, the predictions of instance segmentation results can be divided into different categories, while true positive (*TP*), false positive (*FP*), and false negative (*FN*) represent the number of samples in each category. Then, the corresponding precision value and recall value is described by

$$\text{Precision} = \frac{TP}{TP + FP} \tag{11}$$

$$\text{Recall} = \frac{TP}{TP + FN} \tag{12}$$

With confidence threshold changes, precision and recall will be different, with the result that the precision and recall curve *<sup>P</sup>*(*r*) where the recall value serves as the abscissa and precision value serves as the ordinate in Cartesian coordinate system. Then, the *AP* of a given *IOU* threshold is described by

$$AP\_{IOL} = \int\_0^1 P(r) dr\tag{13}$$

Then, *AP* is the average value of 10 *APIOU* whose *IOU* threshold ranges from 0.5 to 0.95 with the stride of 0.05., which is described by

$$AP = \frac{1}{10} \times \sum\_{IOI=0.50}^{0.95} AP\_{IOI} \tag{14}$$
