SVHighlights

> **Mirror note.** This ModelScope copy additionally includes the raw YouTube videos referenced by `metadata/video_list.csv`, downloaded with yt-dlp (720p mp4) as `videos/<youtube_id>.mp4` (635 of 638 unique IDs). Missing because YouTube reports them unavailable: `107IBonma_o`, `21fJiYJNcik`, `a5_Eqony5RE`. Use `video_map.csv` to map `vid` (e.g. `soccer_5`) to the full-broadcast and highlight YouTube IDs and trim windows. Videos remain subject to their publishers' terms; annotations are CC BY-NC 4.0 from the original authors (research, non-commercial only). Original dataset: https://huggingface.co/datasets/idong1004/SVHighlights # SVHighlights: Towards Extremely Long Sport Video Highlight Detection [](https://arxiv.org/abs/2606.06926) [](https://doi.org/10.1145/3770855.3817564) [](https://github.com/leedongkyu2019/SVHighlights) [](https://leedongkyu2019.github.io/SVHighlights/) [](https://creativecommons.org/licenses/by-nc/4.0/) **Donggyu Lee\*, Youngbin Ki\*, Jeonghun Kang, Taehwan Kim** — UNIST KDD 2026 · Datasets & Benchmarks Track (*equal contribution) <p align="center"> <img src="https://raw.githubusercontent.com/leedongkyu2019/SVHighlights/main/assets/teaser.png" width="60%"> </p> SVHighlights is the first highlight-detection benchmark for **extremely long sports videos** — **320 full-length broadcasts** averaging **2.00 hours** across **8 sports** (40 videos each: american football, baseball, basketball, ice hockey, racing, rugby, soccer, volleyball), totaling **640.18 hours** — roughly 30–60× longer than prior highlight-detection benchmarks. We do **not** redistribute the original videos; per-clip `{0, 1}` highlight labels are distilled by aligning each official highlight video to its full broadcast via pixel-level PSNR matching. This release contains QVHighlights-style features and all preprocessing artefacts needed to reproduce or extend the benchmark. ## 📦 Contents ``` annotations/ alignment.tar # per-video PSNR alignment outputs -> alignment/ segments.tar # context-aware segments (TF-SELECTOR) -> segments/ whisper.tar # word-level WhisperX transcripts -> whisper/ shots.tar # shot boundaries -> shots/ label.json # per-clip {0, 1} highlight labels (eval GT) all_filtered_frame_idx.json # final aligned frame index per highlight clip segment_caption.json # VLM segment captions volume.json, minmax_volume.json features/ <sport>.tar # one tarball per sport (~1.5–2.5 GB each) # extracts to <sport>/{vid_clip,vid_slowfast,txt_clip,aud_pann}/ ``` Directory-style annotations and features are packed as tarballs so the dataset downloads as a few large objects instead of thousands of small files. Video URLs and per-video trim boundaries are released as `video_list.csv` in the [GitHub repo](https://github.com/leedongkyu2019/SVHighlights). ## ⬇️ Download ```bash # 1) Download (use --max-workers 1 to stay under HF's anonymous API rate limit) huggingface-cli download idong1004/SVHighlights \ --repo-type dataset --local-dir ./data --max-workers 1 # 2) Extract the annotation tarballs in place for a in alignment segments whisper shots; do tar -xf data/annotations/${a}.tar -C data/annotations/ && rm data/annotations/${a}.tar done # 3) Extract the per-sport feature tarballs in place for s in american_football baseball basketball ice_hockey race rugby soccer volleyball; do tar -xf data/features/${s}.tar -C data/features/ && rm data/features/${s}.tar done ``` ## 📜 License & video terms The SVHighlights annotations and extracted features are released under **[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)** — free for non-commercial research with attribution. The accompanying code on [GitHub](https://github.com/leedongkyu2019/SVHighlights) is released separately under **Apache 2.0**. The original sports broadcasts are **not** redistributed; `video_list.csv` provides public URLs only, and use of the source videos is subject to the original publishers' terms of service. ## 📚 Citation ```bibtex @inproceedings{Lee_2026, title={SVHighlights: Towards Extremely Long Sport Video Highlight Detection}, url={http://dx.doi.org/10.1145/3770855.3817564}, DOI={10.1145/3770855.3817564}, booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2}, publisher={ACM}, author={Lee, Donggyu and Ki, Youngbin and Kang, Jeonghun and Kim, Taehwan}, year={2026}, month=Aug, pages={9254–9265} } ```