Talk by Lu Xu (Xi’an Jiaotong University)

Speaker: Lu Xu (Xi’an Jiaotong University)

Title: Reducing human supervisions in training 3D hand pose estimators with Bayesian Learning

Abstract: 3D hand pose estimation is one of the critical techniques among several human-computer interface applications. Although deep learning based methods have achieved promising results on 3D hand datasets, the huge amount of works on building kinematic models and annotating hand poses has been the bottleneck of applying these data-hungry algorithms to realistic applications for a long time. This work is trying to answer the question: " how can we alleviate the reliance on human supervisions in training deep models for 3D hand pose estimation"? With the help of Bayesian learning, such as Bayesian probabilistic models and non-parametric Bayesian learning, we present a couple of approaches that allow the network to model the kinematics of hands during training, and learn the regression mapping in a data-efficient way. We show that by being a little Bayesian, deep hand pose estimators can be developed with less human supervisions, and toward realistic applications.