| --- |
| pretty_name: "Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning" |
| license: cc-by-nc-4.0 |
| tags: |
| - arxiv:2610.22444 |
| configs: |
| - config_name: soccerstvg_oracle |
| drop_labels: true |
| data_files: |
| - split: test |
| path: |
| - soccerstvg/videos/oracle/**/*.mp4 |
| - soccerstvg/videos/oracle/metadata.jsonl |
| default: true |
| --- |
| |
| > Upload in progress. Complete configurations: oracle. Pending configurations: context_1min, context_3min, context_5min, context_10min. The table below describes the locally validated full release; only configurations present in the YAML are ready to load. |
|
|
|
|
| # Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning |
|
|
| **Paper:** [arXiv:2610.22444](https://arxiv.org/abs/2610.22444) | **Code:** [THU-KEG/EventSTVG](https://github.com/THU-KEG/EventSTVG) | **Dataset:** [Hugging Face](https://huggingface.co/datasets/zacharyz21/EventSTVG) |
|
|
| EventSTVG evaluates event localization and the spatial grounding of semantic roles. |
| This release contains the frozen SoccerSTVG component: **3,516 distinct target |
| events from 50 soccer matches**, with 17 semantic event types, 16 semantic roles, |
| and 6,497 event–role instances. All five configurations contain the same 3,516 |
| target event IDs, presented with different video context. They are evaluation |
| configurations, not separate training/validation partitions. |
|
|
| Each VideoFolder record integrates its video, complete official prompt, and |
| annotations. No join against an external prompts file is required. The source |
| prompt text, GT answer objects, sample identifiers, and original metadata are |
| preserved unchanged. Original spatial annotations are embedded for frame and |
| pixel-coordinate traceability and compatibility with the frozen evaluator. |
|
|
| ## Configurations |
|
|
| | Config | Split | Records | Input | |
| |---|---|---:|---| |
| | `soccerstvg_oracle` | `test` | 3,516 | Temporally trimmed target event; spatial localization only | |
| | `soccerstvg_context_1min` | `test` | 3,516 | One-minute context | |
| | `soccerstvg_context_3min` | `test` | 3,516 | Three-minute context | |
| | `soccerstvg_context_5min` | `test` | 3,516 | Five-minute context | |
| | `soccerstvg_context_10min` | `test` | 3,516 | Ten-minute context | |
|
|
| Long-context labels describe the benchmark settings; the exact annotated window |
| duration is retained in `metadata.window_duration`. The five settings represent |
| 3,516 unique target events, not 17,580 independent events. The source videos total |
| approximately 1.31 TB across the five settings; streaming avoids downloading the |
| entire collection before inspecting records. |
|
|
| ## Load |
|
|
| The local VideoFolder integration was validated with Python 3.10, Datasets 3.6.0, |
| PyArrow 20.0.0, PyTorch 2.4.1, Torchvision 0.19.1 and PyAV 12.3.0, including native |
| video decoding and streaming for every configuration. Remote loading and viewer |
| results are checked separately after upload. |
|
|
| Install Hugging Face Datasets and a compatible video decoding backend, following |
| the [official video installation instructions](https://huggingface.co/docs/datasets/video_load). |
| For a private repository, authenticate using an account with access (`hf auth login`). |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset( |
| "zacharyz21/EventSTVG", |
| "soccerstvg_context_1min", |
| split="test", |
| streaming=True, |
| token=True, |
| ) |
| record = next(iter(ds)) |
| video = record["video"] |
| prompt = record["prompt"] |
| ground_truth = record["answer"] |
| ``` |
|
|
| To inspect metadata without decoding video, use the genuine Video feature with |
| decoding disabled. This preserves the video association rather than treating a |
| path string as a video column: |
|
|
| ```python |
| from datasets import Video |
| |
| ds = ds.cast_column("video", Video(decode=False)) |
| record = next(iter(ds)) |
| print(record["sample_id"], record["video"]["path"]) |
| print(record["prompt"]) |
| ``` |
|
|
| For local loading after downloading a configuration: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset( |
| "videofolder", |
| data_files={"test": [ |
| "EventSTVG/soccerstvg/videos/oracle/**/*.mp4", |
| "EventSTVG/soccerstvg/videos/oracle/metadata.jsonl", |
| ]}, |
| drop_labels=True, |
| split="test", |
| ) |
| ``` |
|
|
| ## Download selected files |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| snapshot_download( |
| repo_id="zacharyz21/EventSTVG", |
| repo_type="dataset", |
| token=True, |
| local_dir="EventSTVG", |
| allow_patterns=[ |
| "README.md", |
| "soccerstvg/README.md", |
| "soccerstvg/task_spec.json", |
| "soccerstvg/videos/oracle/**", |
| "evaluation/**", |
| ], |
| ) |
| ``` |
|
|
| Replace `oracle` in the allowed path with the desired context setting to download |
| its videos and integrated metadata. Do not download all configurations unless needed. |
|
|
| ## Record fields |
|
|
| | Field | Meaning | |
| |---|---| |
| | `sample_id` | Original stable event identifier, shared across settings | |
| | `file_name` | In `metadata.jsonl`, video path relative to that file's directory; VideoFolder turns this association into `video` | |
| | `video` | Hugging Face `Video` feature produced by VideoFolder | |
| | `prompt` | Complete, unchanged official model-input prompt | |
| | `event_type` | Original final semantic event type, also retained in `metadata` | |
| | `active_roles` | Original requested role list, also retained in `metadata` | |
| | `answer` | Frozen evaluation GT: Oracle localization array, or context object containing `temporal_window` and `role_localization` | |
| | `metadata` | Entire original metadata object: event and role definitions by ID, context setting, source identifiers, crop/window timing, ordering, versions and hashes | |
| | `spatial_annotation` | Entire original spatial annotation: original dimensions, FPS, zero-based keyframe indices, per-role pixel boxes and original tracking records | |
|
|
| The prompt and video are **model inputs**. `answer`, `spatial_annotation`, and |
| target/source timing in `metadata` are **ground truth or provenance** and must not |
| be included in a model request. The source metadata contains original local paths |
| for provenance; use `video` or the relative `file_name` to access released videos. |
|
|
| Hugging Face/Arrow represents heterogeneous role dictionaries as a common struct |
| schema and may show additional null fields for roles not requested by a record. |
| Use `active_roles` as the authoritative requested-role list. The distributed |
| JSONL preserves exactly the original role keys; absent roles are not new labels. |
|
|
| ## Time, frames, boxes, and visibility |
|
|
| - All answer times are in **seconds from the beginning of the input video**, not |
| the on-screen soccer match clock. Oracle time starts at the event clip origin. |
| - Context `answer.temporal_window` gives target start/end within its input video; |
| `role_localization[].time_sec` uses that same input-video origin. |
| - GT support is the explicitly stored nominal 2 Hz sequence: every 0.5 seconds |
| starting at the target start. Do not generate missing terminal samples or |
| interpolate boxes. Oracle supports exactly the timestamp list in its prompt/GT. |
| - Answer boxes are `[xmin, ymin, xmax, ymax]`, normalized to `[0,1000]`, with the |
| origin at the top-left. X coordinates are scaled by video width and Y by height. |
| - An explicit role `null` retains the original benchmark meaning of no reliable |
| localization at that timestamp. It is distinct from missing/invalid prediction |
| fields. A role refers to the same individual throughout the target event. |
| - `metadata.window_start` and `window_end` are source-half-video seconds. |
| `target_start_in_window` and `target_end_in_window` are input-video seconds. |
| `source_half_video`, `game`, `half`, `sn_anchor_time_abs`, and the original |
| occurrence identifiers are preserved in the same record; no separate source |
| manifest is required. |
| - Embedded original spatial `frame_index` is zero-based in the original event |
| clip. Keyframes use `round_half_up(fps_original * 0.5 * k)`. At 25 FPS this is |
| `0,13,25,38,...`; the original actual frame timestamp can be `0.52` where the |
| answer uses nominal `0.5`. Original spatial boxes use original-video pixels, |
| with width/height retained in `spatial_annotation`. |
|
|
| The unchanged official prompt specification and explicit release conventions are |
| in [soccerstvg/task_spec.json](soccerstvg/task_spec.json). |
|
|
| ## Evaluation |
|
|
| The existing frozen `role_stvg_eval_v1.0` evaluator and parser are included in |
| [evaluation/](evaluation/README.md), with their original source modules and policy. |
| The release includes an export utility to reconstruct the evaluator's expected |
| GT layout from integrated metadata; it does not regenerate any labels or prompts. |
| See [evaluation/README.md](evaluation/README.md) for commands and prediction format. |
|
|
| ## Files and video integration |
|
|
| Each setting is a VideoFolder with one `metadata.jsonl` and videos organized by |
| the original sample path. The root YAML configuration includes **both MP4 files |
| and metadata**, with directory-derived classification labels disabled. This is |
| the official [VideoFolder metadata convention](https://huggingface.co/docs/hub/datasets-video), |
| not a JSON-only annotation table. See also the [Datasets VideoFolder guide](https://huggingface.co/docs/datasets/video_dataset) |
| and [Hub configuration guide](https://huggingface.co/docs/hub/datasets-manual-configuration). |
| The JSONL is retained as the release source; any Hub-generated viewer cache is |
| managed by Hugging Face. Viewer availability can depend on processing status and |
| the account's access to private-dataset preview features. |
|
|
| ## Citation |
|
|
| If you use Event-STVG, please cite the paper: |
|
|
| ```bibtex |
| @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 and source |
|
|
| The existing repository license field is retained as **CC BY-NC 4.0**. No new |
| commercial-use permission is granted. The soccer videos originate from SoccerNet |
| match footage; this release does not claim ownership of the upstream footage or |
| replace its applicable access conditions and rights. See the |
| [SoccerNet source website](https://www.soccer-net.org/) for upstream dataset access. |
| Original event identifiers, game/half identifiers, crop timing, and annotations |
| remain traceable in each record. |
|
|
