SoccerNet-FoulRet
SoccerNet-FoulRet: Retrieving Semantically Similar Soccer Foul Videos
A benchmark for semantic foul retrieval in soccer videos with 693 human-verified queries and category-relevance labels to evaluate precedent-based video retrieval.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info soccernet-foulret
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
7 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen3-VL-Emb-2B-FT | 29.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| InternVideo2-1B | 23.1 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-VL-Emb-2B | 23.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| InternVideo2-6B | 22.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-VL-Emb-8B | 20.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| VLM2Vec-V2 | 18.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| LamRA | 18.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Built from the SoccerNet-MVFoul dataset, this benchmark constructs a search gallery and a manual validation interface to create human-verified query-candidate pairs based on refereeing precedent relevance.
It evaluates zero-shot video/vision-language embedders and fine-tuned baselines using HitRate@10. Results show strong zero-shot models achieve under 5% HitRate, highlighting the challenge of transferring category supervision to semantic precedent retrieval.