Paper (arXiv:2610.22444) · PDF · Code · Dataset
Yizhe Zhang, Xujia Wang, Songqi Tang, Xujie Tang, Leo Jiahao Jiang, Jiacheng Wang, Bowen Qu, JinBo Li, Qin Wang, Ji Qi, Lei Hou, Bin Xu, Juanzi Li.
EventSTVG studies event localization and the spatial grounding of semantic roles in video. A query identifies an event and its participants by their roles; a prediction localizes the event in time and tracks the same participant for each requested role.
This first release contains the SoccerSTVG football component: 3,516 target events from 50 matches, 17 event types and 16 role types, forming 6,497 event–role instances. Oracle and 1/3/5/10-minute settings share the same event IDs. Together they form 17,580 evaluation instances, not 17,580 distinct events. The multi-domain extension is pending release.
Dataset: zacharyz21/EventSTVG on Hugging Face. Videos and final annotations are maintained there. Dataset access currently requires requesting approval on the HF page and signing in with an approved account. The paper describes the full benchmark of 4,016 event occurrences across soccer and seven additional domains; the current public release contains the 3,516-event soccer component described above.
Performance on SoccerSTVG, measured by m-vIoU (%).
See baseline protocols and reproduction details and numerical results.
Install with Python 3.10 or later:
python -m pip install -r requirements.txtThe compact evaluator uses only the Python standard library. Matplotlib and pytest support the original report CLI and regression suite. Optional API baseline dependencies are in baselines/requirements-api.txt; local model requirements and access limitations are described in baselines/README.md.
Videos and full annotations belong on Hugging Face. GitHub contains code, documentation and three small annotation/prediction examples per setting, without video files. Each HF setting has one integrated soccerstvg/videos/<setting>/metadata.jsonl, carrying video association, complete prompt and annotations. Read data_format.md before constructing model inputs: annotations and source timing are withheld evaluation information.
Request access on the HF dataset page, then install and sign in. Download metadata only by default:
hf auth login
python scripts/download_data.py --repo-id zacharyz21/EventSTVG \
--setting context_1min --local-dir dataAdd --sample-id 'EXACT_EVENT_ID' to fetch only that event's video. Repeating this flag fetches an explicitly chosen subset. The script never defaults to downloading all videos or all five settings (the existing video collection is approximately 1.31 TB). See data access instructions for revision pinning and local loading.
Run an evaluation example from this directory, without a model or video download:
python evaluation/evaluate.py \
--metadata examples/context_1min.metadata.jsonl \
--predictions examples/predictions/copy_gt/context_1min.jsonl \
--setting context_1min --model copy_gt --out artifacts/quickstart.json
python scripts/verify_examples.pyExpected results for each example are stored in examples/expected_scores.json. copy_gt is a synthetic interface check; it is not a model baseline.
The official metric implementation is role_stvg_eval_v1.0, with frozen parser role_stvg_json_extraction_v1. Roles are matched by identity, thresholds use strict >, and missing/failed predictions remain in the GT denominator. The historical RoleSTVG package names, identifiers and sample IDs are preserved for reproducibility; EventSTVG is the release name. Evaluation documentation includes input contracts, formulas, failure handling and separate commands for all five settings.
Annotation documentation describes the actual human annotation and quality review evidence, with unconfirmed facts listed explicitly. Baseline documentation covers existing API and local implementations, recorded model aliases, media preprocessing, matched 300-event comparisons, and known reproduction gaps.
If you use Event-STVG, please cite the paper:
@misc{zhang2026eventstvgspatiotemporalvideogrounding,
title={Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning},
author={Yizhe Zhang and Xujia Wang and Songqi Tang and Xujie Tang and Leo Jiahao Jiang and Jiacheng Wang and Bowen Qu and JinBo Li and Qin Wang and Ji Qi and Lei Hou and Bin Xu and Juanzi Li},
year={2026},
eprint={2610.22444},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.22444},
}License: the existing data-card label CC BY-NC 4.0 is provisionally retained for newly contributed annotations, subject to confirmation of ownership and scope. SoccerNet match footage and other upstream content retain their own rights and access conditions; this project does not relicense them. Commercial use requires authorization from relevant rights holders. Code license is undecided and is not covered by the data-card license. See LICENSE_SCOPE.md.