Learning Sparse Temporal Video Mapping for Action Quality Assessment in Floor Gymnastics

TL;DR

This paper introduces a new dataset and a novel sparse mapping method for assessing action quality in long-duration sports videos, addressing the limitations of existing short-term focused models.

Contribution

The paper presents a new dataset, AGF-Olympics, for long-term sports analysis and proposes a discriminative attention module for effective long-range spatio-temporal modeling.

Findings

The proposed module effectively captures long-range spatial and temporal correlations.

The dataset includes challenging scenarios with background, viewpoint, and scale variations.

The method outperforms existing short-term models on long-duration gymnastics videos.

Abstract

Athlete performance measurement in sports videos requires modeling long sequences since the entire spatio-temporal progression contributes dominantly to the performance. It is crucial to comprehend local discriminative spatial dependencies and global semantics for accurate evaluation. However, existing benchmark datasets mainly incorporate sports where the performance lasts only a few seconds. Consequently, state-ofthe-art sports quality assessment methods specifically focus on spatial structure. Although they achieve high performance in short-term sports, they are unable to model prolonged video sequences and fail to achieve similar performance in long-term sports. To facilitate such analysis, we introduce a new dataset, coined AGF-Olympics, that incorporates artistic gymnastic floor routines. AFG-Olympics provides highly challenging scenarios with extensive background, viewpoint, and scale variations over an extended sample duration of up to 2 minutes. In addition, we propose a discriminative attention module to map the dense feature space into a sparse representation by disentangling complex associations. Extensive experiments indicate that our proposed module provides an effective way to embed long-range spatial and temporal correlation semantics.