robotwang/MITFLD

MITFLD: MIT Lecture Fragmentation Dataset

MITFLD is a lecture video fragmentation dataset introduced in _Towards Key Point Identification (KPI) for Lecture Videos: Approaches and Performance Evaluation_, serving as a benchmark for lecture video fragmentation methods without using synthetic videos.

It enables research in:

• —Key Point Identification (KPI) in lecture videos
• —Lecture fragment recommendation
• —Non-linear and bite-sized learning
• —Content-based indexing and search for lecture videos

Data Source Attribution

The MITFLD dataset is sourced from [MIT OpenCourseWare](https://ocw.mit.edu/), a free and open publication of material from thousands of MIT courses, enabling global access to high-quality educational resources.

Folder Structure

├── frags
│   └── <video_id>.json         # Ground truth fragment annotations
├── README.md
├── transcripts
│   └── <video_id>.json         # Transcript of the lecture
├── video_id_list.txt           # List of video_ids included
└── videos
    └── <video_id>.mp4          # Lecture video files

KPI Framework

To facilitate experiments, the dataset is designed to work seamlessly with the [kpi](https://bit.ly/kpi_lecture_frag) unified Python framework for lecture video fragmentation, providing:

• —Dataset abstractions
• —Multiple baseline and advanced methods (e.g., BiLSTM, TW-FINCH, PSD)
• —Evaluation metrics (F-score, mMoF, mIoU)
• —Extensibility for your own methods

Citation

If you use this dataset or the kpi framework in your research, please cite:

@article{wang2025towards,
  title={Towards Key Point Identification (KPI) for Lecture Videos: Approaches and Performance Evaluation},
  author={Wang, Jiaqi and Kwok, Ricky Y-K and Ngai, Edith CH},
  journal={ACM Transactions on Multimedia Computing, Communications and Applications},
  year={2025},
  publisher={ACM New York, NY}
}