Synthetic Dataset and Pose Estimation for Mice

Abstract

Pose estimation of RBG-D images is useful in studying the behavior of laboratory mice. We generate synthetic datasets of mice using an anatomical 3D mesh model; we manipulate it and output depth images as heatmaps with ground truth labels of 48 joints corresponding to body parts. Using these datasets, we trained a convolutional neural network to estimate the position of joints given a depth image of a mouse. Using these methods, we achieved 95% accuracy for the head joint and 92% accuracy for other joints.