Uncertainty-aware Score Distribution Learning for Action Quality Assessment

Assessing action quality from videos has attracted growing attention inrecent years. Most existing approaches usually tackle this problem based onregression algorithms, which ignore the intrinsic ambiguity in the score labelscaused by multiple judges or their subjective appraisals. To address thisissue, we propose an uncertainty-aware score distribution learning (USDL)approach for action quality assessment (AQA). Specifically, we regard an actionas an instance associated with a score distribution, which describes theprobability of different evaluated scores. Moreover, under the circumstancewhere fine-grained score labels are available (e.g., difficulty degree of anaction or multiple scores from different judges), we further devise amulti-path uncertainty-aware score distributions learning (MUSDL) method toexplore the disentangled components of a score. We conduct experiments on threeAQA datasets containing various Olympic actions and surgical activities, whereour approaches set new state-of-the-arts under the Spearman's Rank Correlation.