Generating Multiple Diverse Hypotheses for Human 3D Pose Consistent with 2D Joint Detections

摘要

We propose a method to generate multiple diverse and valid human posehypotheses in 3D all consistent with the 2D detection of joints in a monocularRGB image. We use a novel generative model uniform (unbiased) in the space ofanatomically plausible 3D poses. Our model is compositional (produces a pose bycombining parts) and since it is restricted only by anatomical constraints itcan generalize to every plausible human 3D pose. Removing the model biasintrinsically helps to generate more diverse 3D pose hypotheses. We argue thatgenerating multiple pose hypotheses is more reasonable than generating only asingle 3D pose based on the 2D joint detection given the depth ambiguity andthe uncertainty due to occlusion and imperfect 2D joint detection. We hope thatthe idea of generating multiple consistent pose hypotheses can give rise to anew line of future work that has not received much attention in the literature.We used the Human3.6M dataset for empirical evaluation.