Finding the defining moments in a two-hour sports broadcast is very different from analyzing a video that lasts only a few minutes. Yet most benchmarks for AI highlight detection have been built around short-form video, leaving a gap between how these systems are tested and the content they are meant to analyze.

Led by Professor Taehwan Kim of the Graduate School of Artificial Intelligence at UNIST has developed     SVHighlights (Sport Video Highlights) , a benchmark for highlight detection in long sports videos. It contains 320 videos across eight sports, totaling 640.18 hours, with each video averaging two hours—roughly 30 to 60 times longer than those in existing benchmarks.

Building a benchmark at this scale would normally require people to watch hundreds of hours of footage and mark highlights by hand. The researchers instead used official highlight videos produced from the same broadcasts as reference points, developing an algorithm that automatically matches those scenes to their original locations in the full-length video.

The algorithm compares frames at the pixel level while accounting for their order in time—an important distinction in sports broadcasts, where replays can look identical to the original play. Human involvement is largely limited to identifying game boundaries and checking the automatically generated matches. Genuine alignment errors accounted for just 0.18% of frames during manual verification.

The team also developed TF-SELECTOR, a training-free framework for detecting highlights in long videos. Rather than evaluating isolated clips, it groups related shots into coherent segments and uses visual descriptions, transcripts, and audio volume to provide context for a large language model, which assigns each segment a highlight score.  

On SVHighlights, TF-SELECTOR surpassed the second-best baselines by 2.50 points in HIT@1, 4.04 points in HIT@K, and 2.95 points in IoU, measures of how accurately a system identifies and locates highlights.

“By using official highlight videos that broadcasters have already produced, we can replace much of the labor-intensive annotation normally required to build this kind of dataset,” said Professor Kim. “Because the same approach can be used to expand the benchmark, SVHighlights could provide a more realistic basis for evaluating long-form video analysis and developing better highlight-detection methods.”

Donggyu Lee and Youngbin Ki of UNIST contributed equally as first authors. Their work was presented at the 2026 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), held August 9-13 on Jeju Island. The dataset and code are available on the team's project page (https://leedongkyu2019.github.io/SVHighlights/).

The research was supported by the Ministry of Science and ICT (MSIT) through the grants of the Institute of Information & Communications Technology Planning & Evaluation (IITP), including the AI Star Fellowship and Artificial Intelligence Graduate School programs at UNIST.

Journal Reference

Donggyu Lee, Youngbin Ki, Jeonghun Kang, and Taehwan Kim, “SVHighlights: Towards Extremely Long Sport Video Highlight Detection,”     ACM KDD , (2026).