Shooter Moira ; Malleson Charles ; Hilton Adrian

Abstract

Estimating the pose of animals can facilitate the understanding of animalmotion which is fundamental in disciplines such as biomechanics, neuroscience,ethology, robotics and the entertainment industry. Human pose estimation modelshave achieved high performance due to the huge amount of training dataavailable. Achieving the same results for animal pose estimation is challengingdue to the lack of animal pose datasets. To address this problem we introduceSyDog: a synthetic dataset of dogs containing ground truth pose and boundingbox coordinates which was generated using the game engine, Unity. Wedemonstrate that pose estimation models trained on SyDog achieve betterperformance than models trained purely on real data and significantly reducethe need for the labour intensive labelling of images. We release the SyDogdataset as a training and evaluation benchmark for research in animal motion.