Figure 4: Example image with three detections in the context of boat detection shown as bounding boxes. Two boats were correctly de- 
tected (TP, green), and one was incorrectly detected (FP, red). Two boats were overlooked (FN, blue). The recall for this case is 
therefore two out of four (50%) and the precision is two out of three (66%). (Koch et al., 2024) 
Recall (r) as defined in ISO TS 4213 indicates how many objects should have been detected in one 
classification. In Figure 4, there are four vessels. Because only two boats were identified by the model, 
the recall metric is 50%. 
FTP 
Fr =— 
TP+FN 
Another important performance metric for object detectors is Intersection over Union (loU), which 
compares the ground truth bonding boxes with the predicted bounding boxes. The result is a value 
between 0 and 1 and can be set as a threshold in order to determine whether a prediction is a TP or a 
EP, 
= Area of Intersection 
0UÜ=- ee 
Area of Union 
The mAP metric combines precision and recall by calculating the precision-recall curve and averaging 
the precision values for different recall thresholds. It is used in maritime research and industry (Mes- 
3a0ud et al., 2024)(Wang et al., 2024). This precision-recall curve shows how the two values depend 
on each other. The mAP ultimately corresponds to the area under this curve: The Area Under Preci- 
sion-Recall Curve (AUPRC) is one possible performance metric (cf ISO/IEC Technical Specification 4213) 
(6.3.7) (ISO, 2022). 
Semantic segmentation - “mean intersection over union” (mIoU) 
Metrics such as loU are commonly used to measure the performance of a task in various use cases 
such as segmentation, object detection, and object tracking. Different variations such as Probabilistic 
Iintersection over Union (ProbloU), Kalman Filtering Intersection over Union (KFIoU), and Rotating In- 
tersection over Union (RIoU) are discussed in the maritime research (Gao et al., 2024). In this study, 
mMIoU is the metric used for semantic segmentation to evaluate the performance of a model (Huang et 
al., 2024) (Koch et al., 2024). 
in Prasad et al. (2016), it is discussed that the line features in a scene with moving vessels and the 
absence of stationary cues may enable registration only if the vessels in the scene are not rotating. 
Thus, for a general maritime scenario, registration of frames is still a challenge. Strictly speaking, the 
best possible way of dealing with this scenario is the use of the motion and gyro sensors of the ship.