DR‐Net: denoising and reconstruction network for 3D human pose estimation from monocular RGB videos

Abstract

A method is presented for accurately estimating 2D and 3D human poses by simultaneously performing 2D pose denoising and 3D pose reconstruction from noisy 2D human pose sequences. The proposed approach globally modifies the input 2D poses that are locally estimated by recent convolutional neural network-based methods. The denoised 2D poses are efficiently converted into 3D poses in a bottom-up manner using a feed-forward network rather than by optimisation, which is frequently used in existing methods. The proposed denoising and reconstruction network is used with existing 2D human pose estimators to provide state-of-the-art 3D human pose estimation results for large-scale real datasets.