Automatic Face Detection
We found that SealNet’s face detector has precision (the percentage of
predictions that are seal face) of 85.43% and a recall (the percentage
of total seal faces that are correctly predicted) of 86.94% after being
trained on a dataset of 516 photos from one haul-out site on a single
day that contained 1,178 valid seal faces. Figure 3 shows the
accuracy of our model across different threshold levels for predicting a
seal face. As the value of threshold decreases, the precision decreases
to 0 while the recall approaches to 1. On the other hand, if threshold
increases, the precision increases to 1 but the recall will decrease to
0. We chose threshold 0 for our face detector because it gives the best
precision-recall trade-off.
We detected 49 false positives, that is, faces detected by SealNet that
were not faces. Most were caused by vegetation or other parts of the
seal that had face-like shapes (Supplementary Figure 1 ).
SealNet missed on average 43 faces, mostly ones that were angled away
from the camera (false negatives, Supplementary Figure 2 ). We
identified a total 408 unique seals, with an average of 2.9 photos per).
Among these, 74 seals appeared in at least 5 photos.