UNIST researchers introduce SVHighlights, a large-scale benchmark designed to evaluate AI models that extract highlights from long-form sports footage across eight sports, including soccer, baseball, basketball, volleyball, American football, ice hockey, rugby, and racing.

SVHighlights comprises 320 long-form videos totaling 640.18 hours, with each video averaging around two hours to mirror real broadcasts and reduce labeling effort.

Human involvement is limited to marking match start/end points and verifying frame matches, with quality checks showing incorrectly matched frames constituting only 0.18% of total.

Ground-truth labels are automatically generated from official broadcaster highlights, minimizing human labor and enabling scalable evaluation.

A matching algorithm locates official highlight moments within full-game footage at minute/second precision, even when moments are replayed, by comparing frames to ensure accurate temporal alignment.

The dataset uses publicly available official highlights as ground truth to anchor evaluation and address the challenge of precise start and end times.

The project page provides dataset and code access, with support from Korea’s Ministry of Science and ICT and related AI initiatives.

Led by Professor Kim Taehwan, with first authors Lee Donggyu and Ki Youngbin, the work was presented at ACM SIGKDD in Jeju on August 9.

TF-SELECTOR, an AI model developed for long-form highlight extraction, combines scene segmentation, speech recognition, vision-language modeling, and an LLM to assess segment importance from visuals, commentary, and audio cues.

TF-SELECTOR achieved top performance across metrics, outperforming the next-best model by 2.50 percentage points on HIT@1, 4.04 points on HIT@K, and 2.95 points on IoU.

SVHighlights enables objective evaluation of long-form video analysis systems and is poised to spur development of higher-performance models for sports highlight extraction.

Beyond sports, potential applications include film/drama summarization, meeting transcription, and identifying key moments in long-duration surveillance footage.