Abstract

Biologists collect and analyze phenomic (e.g., anatomical or non-genomic) data to discover relationships among species in the Tree of Life. We used computer vision techniques to automate the scoring of characters drawn from the tooth loci of bats using standardized images of bat skulls.

We applied a deformable parts model to detect and localize object parts. To score novel images, it is necessary to first provide annotated training images for each character. The challenge is that the novel, unannotated images may show species that have types, total numbers, and layouts of the teeth that differ from the annotated training images. Our method begins by matching the novel unannotated images to the annotated training images. This allows a transfer of tooth annotations to the unannotated images. We then learn a tree parts model on the transferred annotations, and apply this model to detect and label teeth in the unannotated images. This technique performs well and is close to our upper bound performance by the fully supervised model.

The tool developed from this research is available online at GitHub at github.com/AVATOL/

This matrix, and the tool derived from it, was originally developed using a matrix comprising 8 species with character states illustrated by 20 specimens per species, which is archived as MorphoBank project 1226. After publication and distribution of the tool, we added 16 species, most of which had character states illustrated with 10 specimens per species.

This project was funded by an NSF award (DEB-1208306) to Simmons, Cirranello, and Velazco.