Dense soccer video captioning with global context and player position awareness
TL;DR
The results suggest that global temporal context and explicit player-position cues are useful for generating more detailed soccer commentaries, while also revealing remaining trade-offs in precise timestamp localization.
Abstract
Sports video captioning plays an important role in understanding and describing dynamic visual content, and it is particularly challenging for long soccer broadcasts. Existing dense captioning methods often rely on local clips and may miss match-level temporal context or player-position cues, which can lead to captions with insufficient spatial and event-sequence detail. To address these challenges, we propose Spatio-Contextual Perceiver Resampler (SCPR), a framework for single-anchored dense soccer video captioning. SCPR combines three components: (1) global-local feature integration for incorporating long-range video context around each caption anchor; (2) spatial position encoding based on the absolute 2D locations of players; and (3) cross-modal feature fusion that aligns global, local, and spatial features with a trainable language decoder. Evaluated under the SoccerNet-Caption protocol, SCPR improves several captioning metrics, especially CIDEr, and increases recall and F1 in the spotting stage, while some strict localization and ROUGE-style metrics remain stronger for competing baselines. These results suggest that global temporal context and explicit player-position cues are useful for generating more detailed soccer commentaries, while also revealing remaining trade-offs in precise timestamp localization.