Zero-shot learning and detection of teeth in images of bat skulls

Biologists collect and analyze phenomic (e.g., anatomical or non-genomic) data to discover relationships among species in the Tree of Life. The domain is seeking to modernize this very time-consuming and largely manual process. We have developed an approach to detect and localize object parts in standardized images of bat skulls. This approach has been further developed for unannotated images by leveraging knowledge learned from a few annotated images. The key challenge is that the unlabeled images show bat skulls of 'unknown' species that may have types, total numbers, and layouts of the teeth that differ from the 'known' species appearing in the labeled images. Our method begins by matching the unlabeled images to the labeled ones. This allows a transfer of tooth annotations to the unlabeled images. We then learn a tree parts model on the transferred annotations, and apply this model to detect and label teeth in the unlabeled images. Our evaluation demonstrates good performance, which is close to our upper bound performance by the fully supervised model.