Title

Evaluating the Quality of a Synthesized Motion with the Fréchet Motion Distance

Authors

Antoine Maiorca, Youngwoo Yoon, Thierry Dutoit

Abstract

Motion synthesis is an active research topic in the Deep Learning community. It has various fields of application including character animation, humanoid robots and embodied conversational agents. Designing such algorithm is not a straightforward task and must ensure that the output motion are natural. It is however necessary to assess the quality of a synthesized animation to be able to compare algorithms’ performance. Although that evaluations relying on subjective survey give satisfying estimation on the quality of the animation, gathering such group of people with defined requirements (level of expertise of subjects, e.g.) is expensive, time consuming, and having low reproducibility, which hinders fast development iterations. In this sense, a quantitative evaluation with an objective metric leverages those issues since it does not involve human in the loop. However, the design of such evaluation metric must ensure a strong correlation between it and the human perception on the motion quality. In the computer vision field, the metric called Fréchet Inception Distance (FID) exhibits promising results in assessing synthesized images since it is sensitive to various image artifacts and penalizes the lack of diversity in the generated modalities. This metric has become a standard in the evaluation of image generative models such as Generative Adversarial Networks (GANs). Inspired by the success of FID, we introduce the Fréchet Motion Distance (FMD), a new objective metric to evaluate the quality and
diversity of the synthesized human motions. The code is available at https://github.com/antmaio/FrechetMotionDistance.

Link: https://arxiv.org/pdf/2204.12318.pdf

ISIA Lab publications: https://opendata.umons.ac.be/en/publications/services/html/f105.html