OpenSportsLab/soccernetpro-densedescription-sndvc

SoccerNetPro Dense Description (SNDVC)

SoccerNetPro Dense Description (SNDVC) is a dense video captioning dataset for football matches, converted from SoccerNet caption annotations into a unified half-level JSON format for multimodal video understanding and sequence-to-text tasks.

This release contains train/validation/test split folders together with three annotation files:

• annotations-train.json
• annotations-valid.json
• annotations-test.json

Each annotation item corresponds to one half of one football match and includes metadata, paths to source assets, and temporally grounded English dense captions.

---

Dataset Overview

• Task: dense video captioning
• Format version: 1.0
• Annotation date: 2026-03-21
• Language: English
• Source: SoccerNet caption labels
• Creation pipeline: convert_raw_to_json.py
• Temporal granularity: one item per half video
• Timestamp convention: half-relative timestamps in milliseconds

Split statistics

Split	Matches	Half-level items	Dense captions
Train	281	562	21,388
Valid	92	184	7,306
Test	98	196	7,734
Total	471	942	36,428

Competitions covered

• england_epl
• germany_bundesliga
• italy_serie-a
• spain_laliga
• france_ligue-1
• europe_uefa-champions-league

Note: the test split currently covers five competitions and does not include france_ligue-1.

Seasons covered

• 2014-2015
• 2015-2016
• 2016-2017

---

Repository Structure

.
├── annotations-train.json
├── annotations-valid.json
├── annotations-test.json
├── train/
│   └── <competition>/
│       └── <season>/
│           └── <match_folder>/
│               ├── 1_224p.mkv
│               ├── 2_224p.mkv
├── valid/
│   └── <competition>/
│       └── <season>/
│           └── <match_folder>/
│               ├── 1_224p.mkv
│               ├── 2_224p.mkv
├── test/
│   └── <competition>/
│       └── <season>/
│           └── <match_folder>/
│               ├── 1_224p.mkv
│               ├── 2_224p.mkv
└── README.md

Files inside each match folder

File	Description
1_224p.mkv	First-half video
2_224p.mkv	Second-half video

---

What the Annotation Files Contain

Each of the three annotation files follows the same schema:

• annotations-train.json
• annotations-valid.json
• annotations-test.json

Top-level fields include:

• version
• date
• task
• dataset_name
• metadata
• data

Each entry in data describes a single half-match sample.

Example structure

{
  "version": "1.0",
  "date": "2026-03-21",
  "task": "dense_video_captioning",
  "dataset_name": "OSL-Football-DenseCap (train)",
  "metadata": {
    "source": "SoccerNet caption labels",
    "license": "NDA / research use",
    "created_by": "convert_raw_to_json.py",
    "notes": "Half-relative timestamps; one item per half video; raw Labels-caption.json included in inputs."
  },
  "data": [
    {
      "id": "EE1415_TR_001_H1",
      "metadata": {
        "competition": "england_epl",
        "season": "2014-2015",
        "game": "2015-02-21 - 18-00 Chelsea 1 - 1 Burnley",
        "segment": "first_half",
        "split": "TR",
        "game_index": 1
      },
      "inputs": [
        {
          "type": "video",
          "path": "train/england_epl/2014-2015/2015-02-21 - 18-00 Chelsea 1 - 1 Burnley/1_224p.mkv",
          "fps": 25
        },
        {
          "type": "captions",
          "path": "train/england_epl/2014-2015/2015-02-21 - 18-00 Chelsea 1 - 1 Burnley/Labels-caption.json"
        },
        {
          "type": "features",
          "name": "I3D",
          "path": "features/I3D/england_epl/2014-2015/2015-02-21 - 18-00 Chelsea 1 - 1 Burnley/1_224p.npy",
          "dim": 1024,
          "hop_ms": 160
        }
      ],
      "dense_captions": [
        {
          "position_ms": 816000,
          "lang": "en",
          "text": "Goal! Eden Hazard provides Branislav Ivanovic (Chelsea) with a nice pass inside the box..."
        }
      ]
    }
  ]
}

---

Annotation Semantics

One item = one half

Each annotation item corresponds to exactly one half of one match:

• segment = first_half for half 1
• segment = second_half for half 2

This design avoids concatenating both halves into a single long video and keeps caption timestamps aligned to the original half-level timeline.

Dense captions

The field dense_captions contains a list of temporally grounded English descriptions:

• position_ms: timestamp in milliseconds
• lang: language code (en)
• text: caption text

These captions describe fine-grained football events such as:

• goals
• passes
• fouls
• offsides
• saves
• substitutions
• corners
• free kicks
• cards
• match summaries

Inputs

Each sample lists its multimodal inputs under inputs:

• a video file path
• a raw caption annotation file path
• an I3D feature path with metadata (dim = 1024, hop_ms = 160)

Important: the JSON schema preserves feature references used during preprocessing. The current repository structure is centered on the split folders and annotation files; users should rely on the JSON paths as the source of truth for downstream processing.

---

Identifier Format

Each sample has a compact half-level identifier such as:

• EE1415_TR_001_H1
• EE1415_VA_001_H2
• EE1415_TE_001_H1

The general pattern is:

<LEAGUECODE><SEASONCODE>_<SPLIT>_<INDEX>_H<HALF>

Where:

• <LEAGUECODE> identifies the competition
• <SEASONCODE> identifies the season
• <SPLIT> is one of TR, VA, or TE
• <INDEX> is the match index within the split
• H1 / H2 indicates first or second half

---

Temporal Alignment

This dataset uses half-relative timestamps.

That means:

• timestamps start from the beginning of each half
• no cross-half offset is applied
• position_ms always refers to the local timeline of the half-video item

This convention is explicitly preserved in the dataset metadata:

• one item per half video
• raw Labels-caption.json included in the inputs
• half-relative timestamps retained as-is

This design is convenient for:

• dense video captioning
• temporal grounding
• event-level retrieval
• multimodal sequence modeling

---

Split Semantics

The split code stored in metadata.split is:

• TR = train
• VA = validation
• TE = test

The folder paths inside the JSON are consistent with these split files:

• train items point to train/...
• valid items point to valid/...
• test items point to test/...

---

Current Coverage

This release includes:

• 471 football matches
• 942 half-level samples
• 36,428 dense captions
• 3 seasons
• up to 6 competitions, depending on split

It is suitable for research on:

• dense video captioning
• event-grounded football understanding
• multimodal sequence modeling
• long-form sports video understanding
• video-to-text learning

---

Source Data

This dataset is derived from SoccerNet caption annotations.

Original SoccerNet assets include:

• raw caption labels
• half-level match videos
• precomputed visual features

Please follow SoccerNet access conditions and usage policies when using the underlying data.

---

Licensing

• Conversion / dataset packaging: AGPL-3.0
• Underlying SoccerNet content: NDA / research use

Users are responsible for complying with the original SoccerNet licensing and access restrictions.

---

Citation

If you use this dataset, please cite SoccerNet and reference this converted release.

@dataset{soccernetpro_sndvc,
  title={SoccerNetPro Dense Description (SNDVC)},
  year={2026},
  note={Dense video captioning annotations converted from SoccerNet caption labels}
}

---

Acknowledgements

This dataset builds upon the SoccerNet project and its caption annotations for football video understanding.

---

Contact

For questions about preprocessing, annotation structure, or dataset conversion, please open an issue in the Hugging Face dataset repository.