A Novel Skeleton-based Model with Spine for 3D Human Pose Estimation

Kinematic chain model is widely adopted in 3D human pose estimation tasks while it cannot accurately describe the curvature of the torso. This work proposes for the first time a novel model with spine curve to both express the movement of the limbs and the bending curvature of the torso. We parameterize the spine curve with the Bezier curve as it's controlled by anchor points. In this way, the estimation of the spine curve can be converted back to the estimation of points. The result shows that the anchor points of the spine curve can be estimated accurately by existing networks, similar to other joints. With the help of MOSH++ and SMPL model, the existing datasets can also be employed with our skeleton. We fit the SMPL models with the kinematic chain model and our model respectively, the fitting result shows that our model can provide richer information for the construction of human models.