DenseReg: Fully Convolutional Dense Shape Regression In-the-Wild

  • 2018-02-01 18:53:26
  • Rıza Alp Güler, Natalia Neverova, Iasonas Kokkinos
  • 56

Abstract

In this work, we establish dense correspondences between RGB image and asurface-based representation of the human body, a task we refer to as densehuman pose estimation. We first gather dense correspondences for 50K personsappearing in the COCO dataset by introducing an efficient annotation pipeline.We then use our dataset to train CNN-based systems that deliver densecorrespondence 'in the wild', namely in the presence of background, occlusionsand scale variations. We improve our training set's effectiveness by trainingan 'inpainting' network that can fill in missing groundtruth values and reportclear improvements with respect to the best results that would be achievable inthe past. We experiment with fully-convolutional networks and region-basedmodels and observe a superiority of the latter; we further improve accuracythrough cascading, obtaining a system that delivers highly0accurate results inreal time. Supplementary materials and videos are provided on the project pagehttp://densepose.org

 

Introduction (beta)

None

 

Conclusion (beta)

None