British Machine Vision Conference · 2018 · 84 citations · 0 references
ComputingComputer ScienceComputer VisionMachine LearningRobotics3D ReconstructionDeep LearningForward KinematicsKinematicsPose EstimationRotation RepresentationsHuman Motion AnalysisHuman Pose EstimationKinematic Chain AnalysisKinematic Chain ModelingMotion SynthesisQuaternion RepresentationQuaternionsRotations
Deep learning for 3D human pose sequence prediction and generation is an active area, but existing methods either regress joint rotations—leading to error accumulation and discontinuities—or joint positions—requiring reprojection to maintain skeleton constraints. The study proposes a solution that simultaneously mitigates rotation‑based error accumulation and position‑based reprojection issues. QuaterNet is a recurrent network that encodes rotations as quaternions and applies a forward‑kinematics loss to penalize absolute position errors rather than angle errors. QuaterNet achieves state‑of‑the‑art quantitative performance on short‑term predictions and produces qualitatively realistic long‑term motion comparable to recent graphics‑based neural strategies.
Deep learning for predicting or generating 3D human pose sequences is an active research area. Previous work regresses either joint rotations or joint positions. The former strategy is prone to error accumulation along the kinematic chain, as well as discontinuities when using Euler angle or exponential map parameterizations. The latter requires re-projection onto skeleton constraints to avoid bone stretching and invalid configurations. This work addresses both limitations. Our recurrent network, QuaterNet, represents rotations with quaternions and our loss function performs forward kinematics on a skeleton to penalize absolute position errors instead of angle errors. On short-term predictions, QuaterNet improves the state-of-the-art quantitatively. For long-term generation, our approach is qualitatively judged as realistic as recent neural strategies from the graphics literature.