LVVO Dataset: Lecture Video Visual Objects

The Lecture Video Visual Objects (LVVO) dataset is a benchmark designed for object detection in lecture video frames. It provides high-quality annotations of visual content such as tables, charts, images, and illustrations in real university lecture recordings.

📄 Our Arxiv Paper can be found here: Lecture Video Visual Objects (LVVO) Dataset: A Benchmark for Visual Object Detection in Educational Videos
📥 The dataset can be downloaded here: LVVO Dataset Download

---

📚 Dataset Overview

• Total Images: 4,000 unique frames extracted from lecture videos

• Manually Annotated Subset (LVVO 1k): 1,000 frames

• Automatically Labeled Subset (LVVO 3k): 3,000 frames

• Source: Lecture recordings from videopoints.org, covering 8 instructors across 13 courses and 3 domains (Biology, Computer Science, Geosciences)

• Annotation Tool: VoTT by Microsoft

Each visual object in the images is labeled with one of the following categories:

Category ID	Name
1	Table
2	Chart-Graph
3	Photographic-image
4	Visual-illustration

---

🗂️ Sample Dataset (Mini Version)

To help users quickly explore the dataset structure and format, we provide a sample version containing 10 annotated images.

📦 Full Dataset Download

The complete dataset is hosted on Google Drive and includes three files:

🔗 Download Full LVVO Dataset

File Name	Description
LVVO 1k withCategories.zip	1,000 manually annotated images with categories
LVVO 1k.zip	Same images, single-class annotations
LVVO 3k.zip	3,000 images with automatic bounding boxes

Each version follows this internal structure:

LVVO_x/
├── images/             # Contains all .jpg images
├── labels/             # Contains corresponding annotation files (.json)           
└── dataset_info.json   # Metadata: category names, image ID mappings

---

📝 Annotation Format

Each file in the labels/ folder is a JSON annotation corresponding to an image in images/.

It contains:

• asset: Image metadata

  • name: Image file name (e.g., i116_c425_v7624_i_0146.jpg)
  • image_id: Unique integer identifier
  • size: Dictionary with image width and height in pixels

• objects: A list of annotated visual elements, each containing:

  • class: Category ID
    (1 = Table, 2 = Chart-Graph, 3 = Photographic-image, 4 = Visual-illustration)
  • boundingBox: Bounding box details, including:
    • xmin, ymin: Top-left corner
    • xmax, ymax: Bottom-right corner
    • width, height: Box dimensions in pixels (optional but included)

Note: For the LVVO_3k Automatically Labeled Subset, category information is not available.
All objects are labeled with class = 1, where 1 simply denotes “object” as a general category.

📣 Citation

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

@article{biswas2025lvvo,
title={Lecture Video Visual Objects (LVVO) Dataset: A Benchmark for Visual Object Detection in Educational Videos},
author={Dipayan Biswas and Shishir Shah and Jaspal Subhlok},
journal={arXiv preprint arXiv:2406.00123},
year={2025}
}

---

🔍 Adapted Metadata from External Datasets

This repository includes metadata derived from two external datasets (LDD and LPM), adapted using a consistent filtering pipeline for compatibility with the LVVO dataset.

For details on how these external datasets were processed and which files were retained or excluded, refer to:
📄 meta-files/README.metadata.md

📜 License

This repository includes multiple datasets and metadata files, each under a separate license:

• LVVO Dataset: licensed under CC BY 4.0.
• LDD-derived metadata (meta-files/ldd_filename_selected.csv):
  • Derived from the Lecture Design Dataset, licensed under CC BY 4.0.
• LPM-derived metadata (meta-files/lpm_filename_selected.csv):
  • Derived from the LPM Dataset, licensed under CC BY-NC-SA 4.0.

See LICENSE.txt for full details.

---

For questions or additional information, please contact the author at dipayan1109033@gmail.com.