metadata

license: fair-noncommercial-research-license
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
dataset_info:
  features:
    - name: video_path
      dtype: string
    - name: intent
      dtype: string
  splits:
    - name: train
      num_bytes: 56698
      num_examples: 1051
    - name: validation
      num_bytes: 14792
      num_examples: 265
  download_size: 23167
  dataset_size: 71490

Basketball Video-Text Dataset

Overview

A curated subset of the NSVA dataset designed to establish benchmarks for sports video-text models. This dataset addresses the gap in sports-specific evaluation metrics for video understanding models.

Dataset Structure

Train: 1,051 video-text pairs
Validation: 265 video-text pairs
Format: Files with video_path and intent columns
Domain: Basketball action descriptions

Why This Dataset?

Existing video-text datasets focus on abstract-level information but lack sports-specific details. For example, we don't need models to predict generic descriptions like "basketball players are playing basketball" - we need them to understand specific actions like "Three-point 26' jump-shot missed and rebound."

As shown in the NSVA paper, their approach provides compact, actionable information for statistics counting and game analysis, unlike existing datasets that offer vague descriptions. With the rise of video understanding models, we need standardized benchmarks for sports video analysis. This subset enables practitioners to train and validate their models with consistent evaluation metrics.

Usage

This dataset contains video filenames and corresponding action descriptions. To access the actual video files, follow the instructions and use the tools provided in the original NSVA repository: https://github.com/jackwu502/NSVA

Acknowledgement

@inproceedings{dew2022sports,
  title={Sports Video Analysis on Large-Scale Data},
  author={Wu, Dekun and Zhao, He and Bao, Xingce and Wildes, Richard P.},
  booktitle={ECCV},
  month = {Oct.},
  year={2022}
}

License

By using this dataset, you must ensure your use falls under fair use as defined by the original authors. All credits go to the original NSVA dataset creators.