The Winner-Takes-All (WTA) training paradigm inherently prioritises global trajectory accuracy. For long-tail samples, where terminal prediction is inherently difficult, this prioritisation leads to degraded Final Displacement Error (FDE). This asymmetry is reflected in the considerably poorer performance of FDE than Average Displacement Error (ADE) on such samples, yet its underlying mechanisms remain insufficiently diagnosed in the literature. The core contribution of this paper is a two-stage diagnostic framework that reveals two deeper asymmetries and identifies a previously unrecognised phenomenon, which we term hollow diversity. The framework is designed to be independent of specific model architectures and datasets. We first deployed the framework on the JAAD dataset using a custom-implemented baseline and subsequently validated its effectiveness on the ETH/UCY datasets.
Zimo Zhuang (Wed,) studied this question.