Establishing a Unified Evaluation Framework for Human Motion Generation: A Comparative Analysis of Metrics

Evaluating generative models is one of the most challenging tasks to achieve (Naeem et al., 2020). This kind of challenge is largely absent in discriminative models, where evaluation primarily involves comparison with ground truth data. However, for generative models, evaluation involves quantifying the validity between real samples and those generated by the model. A common method for evaluating generative models is through human judgment metrics, such as Mean Opinion Scores (MOS) (Streijl et al., 2016). However, this type of evaluation assumes a uniform perception among users regarding what constitutes ideal and realistic generation, which is often not the case. For this reason, generative models require quantitative evaluation based on measures of validity between real and generated samples. This similarity is quantified on two dimensions: fidelity and diversity. On the one hand, fidelity is the measure of similarity between real and generated spaces on the marginal distribution scale. On the other hand, diversity is the measure of how varied a set of samples is, indicating the extent to which the diversity of the generated set in generative models aligns with the diversity of the real set.