Optimize quaternion slerp inference implementation

Optimize the implementation of quaternion slerp-based pose hypothesis generation to eliminate or reduce the approximately 10–20 millisecond inference-time overhead caused primarily by repeated conversions between SMPL’s axis-angle representation and quaternions.

Background

The paper compares inference speed across axis-angle, Euler-angle, and quaternion representations for probabilistic human pose generation. Although quaternion flows with spherical linear interpolation (slerp) achieve the lowest pose error, they require approximately 10–20 milliseconds more processing time than linear quaternion interpolation and other common rotation representations.

The authors attribute this computational bottleneck mainly to conversions between SMPL’s native axis-angle representation and quaternion representation. They explicitly identify developing a more optimized implementation as an unresolved challenge, making this a concrete open engineering problem associated with the proposed CQF-HMR method.