2.3.1 Object detection evaluation
We evaluated the OD against the movement data (manually annotated and groundtruthed) described in section 2.2 and calculated precision, recall and F1. The precision is the rate of true positives relative to total detections and the recall is the rate of detection of true positives. We used the F1 score (the harmonic mean of the precision and recall) to assess the performance of our object detection model in answering ecological questions on abundance.
\begin{equation} \left(1\right)\ \ \ Precision=\frac{\text{TruePositives}}{TruePositives+FalsePositives}\nonumber \\ \end{equation}\begin{equation} \left(2\right)\ \ \ \ Recall=\frac{\text{TruePositives}}{TruePositives+FalseNegatives}\nonumber \\ \end{equation}\begin{equation} \left(3\right)\ \ \ \ F1=2*\ \frac{precision*recall}{precision+recall}\nonumber \\ \end{equation}
Additionally, we determined the model’s ability to fit a segmentation mask around the fish through the mean average precision value (mAP) (Everingham et al., 2010). We used the mAP50 value, which is the model’s capacity to overlap a segmentation mask around 50% of the ground-truth outline of the fish. A high mAP50 value means that the model has high accuracy when overlapping a mask around the fish. We used the COCO evaluation python script to calculate mAP50 (Massa & Girshick, 2018).