UNIST develops AI benchmark for sports video highlights

Researchers at UNIST have developed ‘SVHighlights’ (Sport Video Highlights), a large-scale benchmark designed to evaluate how accurately artificial intelligence can identify decisive moments in sports broadcasts. Unlike previous benchmarks that relied on short clips of 2 to 4 minutes, SVHighlights utilizes 320 videos across eight sports—including soccer, baseball, basketball, and racing—totaling over 640 hours. The average video length is approximately two hours, making it 30 to 60 times longer than existing datasets and more reflective of real-world broadcasting environments.

To create the benchmark without the high cost of manual labeling, the team developed a matching algorithm that uses official broadcaster highlights as the ‘ground truth.’ The algorithm compares pixels and temporal sequences to ensure that actual game footage is matched rather than repeated replay clips.

Alongside the benchmark, the team introduced ‘TF-SELECTOR,’ an AI model that combines scene segmentation, speech recognition, vision-language models, and large language models (LLMs) to extract highlights. In testing, TF-SELECTOR outperformed existing models in key metrics, including HIT@1 and HIT@K. The researchers noted that this technology could eventually be applied to summarizing movies, organizing meeting records, and searching for key moments in long-term surveillance footage.