Deep Refinement Convolutional Networks for Human Pose Estimation

This work introduces a novel Convolutional Networkarchitecture (ConvNet) for the task of human poseestimation, that is the localization of body joints in a singlestatic image. The proposed coarse to fine architecture addressesshortcomings of the baseline architecture that stem from thefact that large inaccuracies of its coarse ConvNet cannot becorrected by the refinement ConvNet that refines the estimationwithin small windows of the coarse prediction. This is achievedby a) changes in architectural parameters that both increase theaccuracy of the coarse model and make the refinement modelmore capable of correcting the errors of the coarse model,b) the introduction of a Markov Random Field (MRF)-basedspatial model network between the coarse and the refinementmodel that introduces geometric constraints and c) a trainingscheme that adapts the data augmentation and the learning rateaccording to the difficulty of the data examples. The proposedarchitecture is trained in an end-to-end fashion. Experimentalresults show that the proposed method improves the baselinemodel and provides state of the art results on the FashionPose[8] and MPII benchmarks [1].