Machine-assisted photo identification processes require significant amounts of data for each member of a population of interest but offer the possibility to alleviate a significant amount of manual effort. Gathering such data is time consuming and opportunistic, leading to imbalanced datasets ill-suited for traditional machine (deep) learning efforts. Incomplete classifiers, trained on a subset of classes in a population, can be initially useful to identify the most commonly seen individuals. This study investigates the use of incomplete classifiers trained on a subset of often-observed individual killer whales to generate latent space representations of the larger population containing unseen individuals. These semantically relevant representations are subsequently clustered to investigate the efficacy of this method as a secondary identification mechanism. This method proves to be robust to a significant amount of noise while being able to isolate individuals unknown to the classifier when applying limited expert knowledge to the approximate size of the population.