VISTA Dataset

Dataset Structure

dataset/
├── videos/                 # Video files directory
│   ├── train_part1/        # Training set videos (first 8000 samples)
│   ├── train_part2/        # Training set videos (remaining samples)
│   ├── val/                # Validation set videos
│   └── test/               # Test set videos
├── train_part1.json        # Training set metadata (first 8000 samples)
├── train_part2.json        # Training set metadata (remaining samples)
├── val.json                # Validation set metadata
├── test.json               # Test set metadata
└── dataset_info.json       # Dataset description and statistics

Note: Due to the limitation of Hugging Face datasets where a single file folder can store up to 10000 files, the training set videos have been split into train_part1 and train_part2 folders.

Video Files

The video files are placed in the videos/ folder. Videos are organized by split (train_part1/train_part2/val/test).

Metadata Format

Each sample includes the following information:

• id: Sample ID
• title: Paper title
• authors: List of authors
• abstract: Paper abstract (ground truth summary)
• video_file: Video filename
• video_path: Path to the video file (this is processed as a video feature)
• paper_url: Link to the paper PDF
• venue: Publication venue

Usage with Hugging Face Datasets

from datasets import load_dataset

# Load the dataset from Hugging Face Hub
dataset = load_dataset("dongqi-me/VISTA")

# Access splits
train_part1 = dataset["train_part1"]
train_part2 = dataset["train_part2"]
val_data = dataset["validation"]
test_data = dataset["test"]

# Access videos
video_part1 = train_part1[0]["video_path"]  # Access video in train_part1
video_part2 = train_part2[0]["video_path"]  # Access video in train_part2

# Or load from local directory
local_dataset = load_dataset("json", data_files={
    "train_part1": "train_part1.json",
    "train_part2": "train_part2.json",
    "validation": "val.json",
    "test": "test.json"
})

Citation

If you use this dataset in your research, please cite:

@misc{liu2025talk,
      title={What Is That Talk About? A Video-to-Text Summarization Dataset for Scientific Presentations}, 
      author={Dongqi Liu and Chenxi Whitehouse and Xi Yu and Louis Mahon and Rohit Saxena and Zheng Zhao and Yifu Qiu and Mirella Lapata and Vera Demberg},
      year={2025},
      eprint={2502.08279},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.08279}, 
}