qhjqhj00 frechet-motion-distance
Evaluating the Quality of a Synthesized Motion with the Fr'echet Motion Distance — Maiorca et al. (2022) (arXiv:2204.12318, 2022)
What this evaluates
Evaluates the quality and diversity of synthesized human motion by measuring the distributional distance between ground truth and synthetic motion sequences in a learned latent space.
Datasets
- Human3.6M — total ?; splits: train (-1), test (-1); repo https://github.com/antmaio/FrechetMotionDistance
Metrics
Fréchet Motion Distance (FMD)(primary) — range: [0, ∞)- Computes the Fréchet distance between the latent distributions of clean ground-truth motion and synthetic/noisy motion, both encoded via a ResNet34-based autoencoder. Assumes each distribution is approximated by a multivariate Gaussian.
Input / output format
Input: Motion sequences represented as 3D images (x, y, z coordinates as RGB channels) or their corresponding latent vectors from the autoencoder.
Output: A single scalar value representing the Fréchet distance between the two motion distributions.
Scoring recipe
z_clean = autoencoder.encode(clean_motion)
z_noisy = autoencoder.encode(noisy_motion)
mu_c, Sigma_c = estimate_gaussian(z_clean)
mu_n, Sigma_n = estimate_gaussian(z_noisy)
diff = mu_c - mu_n
covmean = sqrtm(Sigma_c @ Sigma_n)
fmd = diff @ diff + trace(Sigma_c + Sigma_n - 2 * covmean)
return fmd
Common pitfalls
- FMD shows reduced sensitivity to temporal noise/discontinuities compared to spatial noise, which may underestimate degradation in time-varying artifacts.
- The metric assumes latent distributions can be well-approximated by Gaussians; non-Gaussian latent structures may yield misleading distances.
- Motion length variations (e.g., 18 vs 64 frames) can affect latent encoding and FMD scores if not normalized or handled consistently.
Evidence (verbatim from paper)
To validate our proposed metric, we need to find that it correctly measures the intensity of motion degradation. The test set is manually altered by adding noise samples with a fixed intensity factor $\zeta$ so that it can play the role of a synthetic motion dataset with artifacts. Then, FMD score is computed between the clean test set and the altered one.
Citation
@misc{maiorca2022frechetmotiondistance,
title={Evaluating the Quality of a Synthesized Motion with the Fr'echet Motion Distance},
author={Maiorca et al. (2022)},
year={2022},
note={arXiv:2204.12318}
}
- arXiv: 2204.12318