Generative Multi-View Based 3D Human Pose Estimation

Large amounts of annotated data is essential for modern Human pose estimation. We propose using a semi supervised learning scheme to estimate the 3D poses from adversarial multi-views generated representations from a single RGB image. Our GAN generated views are the result of training that aims to create authentic and less degenerated outputs. Our method targets the shared latent space between the 3 dimensional and 2 dimensional poses and aims to simplify it by constraining the latent distribution. This resulted in a noticeable increase in the method generalization and exploitation of unlabeled depth maps. We utilized heatmaps to visualize the attention robustness under variety of poses. Our method competes with state of the art performances among semi supervised approaches and excels in some challenging poses as evaluated on Human3.6M, MPII-INF-3DHP and Leeds SportsPose challenging datasets. (1)