Accuracy in Seal Identification
Our closed set data contained 74 seals that had at least 5 photos (607
photos in total). For each fold, the testing set contains one-fifth of
the number of photos of each of the 74 seals and the training set will
contain the remaining photos of those seals. We trained and tested both
PrimNet and SealNet on the same data for each fold. Our average rank-1
accuracy was 88% and our average rank-5 accuracy was 96% across 5
-folds (Figure 5 ). PrimNet yielded 70% rank-1 accuracy and
91% rank-5 accuracy on the same dataset.
Our open set data also included 74 seals with at least 5 photos and 571
photos from seals with fewer than 5 photos. Both PrimNet and SealNet
models were trained and tested utilizing the same splits of data and
equivalent parameters for number of epochs and batches per epoch to
ensure fairness. SealNet outperformed PrimNet with an accuracy that is
9% better at correctly classifying Rank 5 and Rank 1 matches for
probes, respectively (Table 2 ). F1 scores, a measure of model
performance for unbalanced datasets, showed a similar result with
SealNet performing 37% and 39% better than PrimNet.