Convolutional Mesh Regression for Single-Image Human Shape Reconstruction

This paper addresses the problem of 3D human pose and shape estimation from asingle image. Previous approaches consider a parametric model of the humanbody, SMPL, and attempt to regress the model parameters that give rise to amesh consistent with image evidence. This parameter regression has been a verychallenging task, with model-based approaches underperforming compared tononparametric solutions in terms of pose estimation. In our work, we propose torelax this heavy reliance on the model's parameter space. We still retain thetopology of the SMPL template mesh, but instead of predicting model parameters,we directly regress the 3D location of the mesh vertices. This is a heavy taskfor a typical network, but our key insight is that the regression becomessignificantly easier using a Graph-CNN. This architecture allows us toexplicitly encode the template mesh structure within the network and leveragethe spatial locality the mesh has to offer. Image-based features are attachedto the mesh vertices and the Graph-CNN is responsible to process them on themesh structure, while the regression target for each vertex is its 3D location.Having recovered the complete 3D geometry of the mesh, if we still require aspecific model parametrization, this can be reliably regressed from thevertices locations. We demonstrate the flexibility and the effectiveness of ourproposed graph-based mesh regression by attaching different types of featureson the mesh vertices. In all cases, we outperform the comparable baselinesrelying on model parameter regression, while we also achieve state-of-the-artresults among model-based pose estimation approaches.