qhjqhj00 beatv2-eval
SentiAvatar: Towards Expressive and Interactive Digital Humans — Chuhao Jin et al. (2026) (arXiv:2604.02908, 2026)
What this evaluates
Evaluates cross-dataset generalization of co-speech gesture generation on a standard English benchmark, measuring gesture quality, beat consistency, and diversity.
Datasets
- BEATv2 — total ?; splits: test (-1)
Metrics
FGD(primary) — range: other- Fréchet Gesture Distance measuring gesture quality distribution match. Lower is better.
BC— range: [0, 1]- Beat Consistency measuring synchronization with audio beats. Higher is better.
Diversity— range: other- Mean pairwise L2 distance among motion latent features. Higher is better.
Input / output format
Input: Audio and text conditioning.
Output: Generated 3D co-speech gesture sequences.
Scoring recipe
FGD: Compute mean and covariance of real vs generated gesture latent features; FGD = ||μ_r - μ_g||^2 + Tr(Σ_r + Σ_g - 2(Σ_r Σ_g)^0.5).
BC: Measure consistency of generated gestures with audio beat events.
Diversity: Compute mean pairwise L2 distance among generated gesture latent features.
Common pitfalls
- FGD and BC are standard BEATv2 metrics; lower FGD and higher BC indicate better performance.
- Diversity is computed on BEATv2's latent space, not raw coordinates.
- Model is retrained on BEATv2 before evaluation to verify cross-dataset generalization.
Evidence (verbatim from paper)
On BEATv2, we follow the standard protocol and report FGD ($\downarrow$, Fréchet Gesture Distance), BC ($\uparrow$, Beat Consistency), and Diversity.
Citation
@misc{jin2026sentiavatar,
title={SentiAvatar: Towards Expressive and Interactive Digital Humans},
author={Chuhao Jin et al. (2026)},
year={2026},
note={arXiv:2604.02908}
}
- arXiv: 2604.02908