Zero-shot multimodal model spots events in full soccer broadcasts
Original title: Zero-Shot Multimodal Event Spotting in Soccer Broadcasts
On three annotated matches with 130 events, the best setup reaches 72.0% F1@5s and cuts inference cost 72% versus frames.
The paper presents a zero-shot framework that detects, classifies, and temporally localizes events in full-length soccer broadcasts without task-specific parameter training. Overlapping clips are represented by a visual storyboard plus aligned commentary and processed by a general-purpose multimodal model. On three manually annotated matches with 130 events, the best configuration reaches 72.0% F1@5s and beats SoccerNet-trained CALF and NetVLAD++ in a cross-dataset setting. An evidence-grounded excitement-scoring stage ranks events with natural-language explanations for highlight selection.
- No task-specific parameter training is used.
- Storyboards plus commentary replace individual-frame inference.
- Storyboards reduce inference cost by 72% versus frames.
- The method outperforms CALF and NetVLAD++ cross-dataset.
- An excitement score ranks events with written explanations.
From the original article
1 min read at arxiv.orgAutomatic soccer event spotting supports highlight generation, video indexing, and sports analytics, but existing methods depend on large annotated datasets and soccer-specific training. We present a zero-shot framework that detects, classifies, and temporally localizes events in full-length soccer broadcasts without task-specific parameter training.
By Sadia Kamal · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.