Reproduce “QuaterNet: A Quaternion-based Recurrent Model for Human Motion” on OpenResearch

About this paper

Dario Pavllo, David Grangier, Michael Auli

QuaterNet introduces a recurrent neural network model that integrates quaternion representation for joint rotations and a differentiable forward kinematics loss, improving 3D human pose prediction and long-term motion generation. The model achieves state-of-the-art results for short-term prediction on the Human3.6M dataset and provides precise artist control over realistic long-term motion.

Key insights

• Quaternions provide a numerically stable and continuous representation for joint rotations within deep learning models, avoiding issues like gimbal lock inherent in Euler angles.
• Optimizing directly for 3D joint positions via a differentiable forward kinematics loss leads to improved accuracy and training stability compared to optimizing for angle errors, while preserving kinematic constraints.
• Scheduled sampling is essential for preventing catastrophic error accumulation and ensuring stable long-term prediction in autoregressive motion models.

Summary by alphaXiv.