Grouped Guidance Convolutional Networks for Absolute 3D Human Pose and Joint Rotations Estimation
Hua Zhang, Shujie Li, Wei Jia · 2024
In the current landscape of monocular single-person 3D human pose estimation, the primary focus lies in recovering the 3D coordinates of human joints relative to the root joint from monocular images or videos. However, there has been a disregard for the estimation of absolute 3D pose and joint rotations, limiting the application of human pose estimation in fields such as virtual reality and computer animation. In this paper, we address this issue by proposing a Grouped Guidance Convolutional Network capable of estimating 3D absolute pose and joint rotations concurrently. The input to the network consists of 2D joint coordinates, preprocessed using camera parameters. The Grouped Guidance Network extracts shared features between 3D joint coordinates and joint rotations, and through training, the network obtain joint rotations, while a Skeleton Length Network estimates the lengths of human bones. The joint rotations and skeleton lengths are then fed into a forward kinematics layer to obtain the root-relative pose. Simultaneously, the preprocessed data, along with foot contact information outputted by the Foot Contact Network, is input into a Trajectory Network. This network extracts potential connections between foot contact information and the trajectory of the human root joint, further outputting the coordinates of the root joint trajectory. Finally, by combining the root-relative pose and the root joint trajectory, the 3D absolute pose of the human body is obtained. Quantitative and qualitative experiments demonstrate that the proposed method achieves optimal results in absolute pose estimation, surpassing other methods capable of outputting joint rotations.