COMPUTER VISION APPROACHES IN PALEONTOLOGY: A TALE OF TWO METHODS (Invited Presentation)

In the past ten years, morphometrics has been increasingly used to characterize morphological variation in complex biological structures in paleontological contexts. The general approach has been to image specimens at high resolution and to later digitize several anatomical landmarks of interest in each specimen. These anatomical landmarks are then used in multivariate shape analyses, allowing researchers to test specific hypothesis regarding the ecology or evolution of complex phenotypes.

While the quality of imaging techniques and the density of landmarks has continuously increased during the last decade, landmark data is still predominantly collected through manual digitization. Manual digitization of landmarks is, however, not only time and labor-intensive, but also subject to significant amounts of inter-observer bias, precluding datasets from different studies from being confidently combined.

Recent advances in the field of computer vision provide a compelling alternative to manual digitization. In this talk, I will describe two high-throughput automated landmarking methods to collect high-dimensional morphometric data in paleontological contexts. These two methods differ in one key aspect. One is specifically designed to deal with two-dimensional (2D) images, and the other to deal with three-dimensional (3D) surfaces. These computer vision approaches were developed to allow for rapid and dense phenotyping with no significant impact on specimens. Given the increasing adoption of morphometric approaches by the paleontology community, I expect these methods to have broad appeal and to provide a pathway for integration of paleontological and neontological data. Both software presented here are open source.