← ArchivePaper2023

Objective Evaluation Metric for Motion Generative Models: Validating Frechet Motion Distance

Antoine Maiorca, Hugo Bohy, Youngwoo Yoon, Thierry Dutoit

Validates Frechet Motion Distance as an objective metric for evaluating motion generative models, addressing artifacts like foot skating.

Open the paper ↗Trace in the lineage graph

Abstract

Proposes Fréchet Motion Distance (FMD), an objective metric for evaluating motion-generative models that validates performance on realistic motion artifacts like foot skating and over-smoothing. Uses a Transformer-based autoencoder as feature extractor and demonstrates robustness to motion length variations, with significant correlation to human subjective ratings on gesture motion datasets.

How to read this

Category: Evaluation metric: objective measure for motion generative models
Contributions: Frechet Motion Distance (FMD), an objective metric for evaluating motion-generative models
A Transformer-based autoencoder as the feature extractor, with robustness to varying motion length
Validation showing sensitivity to artifacts like foot skating and over-smoothing, and significant correlation with human subjective ratings on gesture datasets
Context: Supports the evaluation of text-to-motion and motion-generation work such as HumanML3D (Guo 2022), adapting Frechet-distance-style metrics to human motion.Builds on: Generating Diverse and Natural 3D Human Motions from Text
Correctness: Correlation with human ratings is shown on gesture motion datasets; the metric inherits the feature extractor's biases, so reported validity should not be over-generalized to all motion domains.
Clarity: Accessible; a single careful pass conveys the metric and its validation, with a second pass only for the autoencoder details.
How to read it: First pass to understand what FMD measures and how it was validated against human judgment; consult details only if adopting it to score your own motion models.

Builds on

Generating Diverse and Natural 3D Human Motions from Text 2022

Built upon by

Nothing yet.

Related work

Dog Code: Human to Quadruped Embodiment Using Shared Codebooks 2024 / MIG
Physically Based Motion Transformation 1999 / SIGGRAPH
Character Motion Synthesis by Topology Coordinates 2009 / CGF
Neural Animation Layering for Synthesizing Martial Arts Movements 2021 / SIGGRAPH

Keywords

fréchet motion distance motion generative models autoencoder transformer foot skating over-smoothing motion synthesis evaluation metric mocap genea challenge

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →