Abstract
We present KAMA, a 3D Keypoint Aware Mesh Articulation approach that allowsus to estimate a human body mesh from the positions of 3D body keypoints. Tothis end, we learn to estimate 3D positions of 26 body keypoints and propose ananalytical solution to articulate a parametric body model, SMPL, via a set ofstraightforward geometric transformations. Since keypoint estimation directlyrelies on image clues, our approach offers significantly better alignment toimage content when compared to state-of-the-art approaches. Our proposedapproach does not require any paired mesh annotations and is able to achievestate-of-the-art mesh fittings through 3D keypoint regression only. Results onthe challenging 3DPW and Human3.6M demonstrate that our approach yieldsstate-of-the-art body mesh fittings.