Self-supervised Secondary Landmark Detection via 3D Representation Learning

Recent technological developments have lead to great advances in the computerized tracking of joints and other landmarks moving animals, including humans. Such promises important biology biomedicine. Modern models depend critically on labor-intensive annotated datasets primary by non-expert However, such annotation approaches can be costly impractical for secondary landmarks, that is, ones reflect fine-grained geometry are often specific customized behavioral tasks. Due visual geometric ambiguity, non-experts not qualified landmark annotation, which require anatomical zoological knowledge. These barriers significantly impede downstream studies because learned exhibit limited generalizability. We hypothesize there exists a shared representation between range motion approximately spanned landmarks. present method learn this spatial relationship three dimensional space, can, turn, self-supervise detector. This 3D learning is generic, therefore applied various multiview settings across diverse organisms, macaques, flies,