ashimdahal/povqa-preprocessed

POVQA Preprocessed Frames and Subtitle-Frame Alignments

Accepted to the MAR Workshop at CVPR 2026.

Paper | Preprocessed Dataset | Project Website

This dataset release contains the **preprocessed artifacts** used by POVQA. It is intended to support research reproducibility without redistributing raw source videos or raw subtitle files.

What is included

• Preprocessed frame bundles for 12 movies and 4 pooling variants
• KEY_FRAMES bundles where available
• Sanitized subtitle-to-frame alignment JSON files
• Per-movie run_summary.json metadata
• A manifest.json file describing the staged release

Release Notes

• Total staged size: approximately 5.30 GB
• Total staged files: 133
• Selected source size before packaging: 4.73 GB
• Selected source file count before packaging: 341471

Directory Layout

Each movie has its own subdirectory:

<movie>/
  metadata_text_centric.json
  metadata_text_centric_blend_blur_with_last_frame.json
  metadata_text_centric_weighted_average.json
  metadata_text_centric_weighted_average_exponential.json
  metadata_text_centric_weighted_average_ramp.json
  run_summary.json
  KEY_FRAMES.tar
  blend_blur_with_last_frame.tar
  weighted_average.tar
  weighted_average_exponential.tar
  weighted_average_ramp.tar

The .tar archives preserve the original folder structure of the preprocessed release and can be extracted with standard tooling.

Notes on Data Content

• This release contains derived, preprocessed artifacts only.
• Raw videos and raw subtitle files are not included in this dataset repo.
• Subtitle credit / torrent source watermark text was scrubbed from the published metadata JSON files while preserving timing and frame-mapping fields.

Intended Use

This release is intended for:

• reproducibility of POVQA preprocessing outputs
• inspection of frame-selection / pooling outputs
• research on video-question-answering pipelines that consume derived frame data

Responsible Use

Users are responsible for ensuring that their use of this dataset complies with applicable law, platform terms, and any rights associated with the underlying source media in their jurisdiction.

Citation

If you use this dataset, please cite the POVQA paper:

@article{dahal2025povqa,
  title   = {POVQA: Preference-Optimized Video Question Answering with Rationales for Data Efficiency},
  author  = {Dahal, Ashim and Ghimire, Ankit and Murad, Saydul Akbar and Rahimi, Nick},
  journal = {arXiv preprint arXiv:2510.01009},
  year    = {2025},
  url     = {https://arxiv.org/abs/2510.01009}
}