Motion In-Betweening via Deep ∆-Interpolator.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human motion synthesis is crucial for animation, robotics, and virtual reality.
• Existing methods often struggle with accuracy and robustness, particularly when conditioned on sparse key frames.

Purpose of the Study:

• To develop a more accurate and effective deep learning-based method for synthesizing human motion from key frames.
• To investigate the benefits of operating in a 'delta mode' for motion interpolation.

Main Methods:

• A deep learning interpolator was designed to operate in the delta mode.
• The spherical linear interpolator (SLERP) was used as a baseline for comparison.
• Experiments were conducted on publicly available human motion datasets.

Main Results:

• The delta mode interpolator achieved state-of-the-art performance on benchmark datasets.
• The delta-regime was shown to be effective when referenced to the last known frame (zero-velocity model).

Conclusions:

• Operating in a local reference frame (delta mode) is more accurate and robust than using a global reference frame for motion synthesis.
• The proposed deep learning approach offers significant improvements in human motion synthesis accuracy and effectiveness.