README.md

---
license: mit
task_categories:
  - visual-document-retrieval
tags:
  - real-data
  - lecture-slides
  - document-analysis
---

# RealSlide: Benchmark for Lecture Slide Analysis

This repository contains the RealSlide benchmark dataset, a collection of real lecture slides curated to evaluate models for slide element detection and text query-based slide retrieval. The dataset complements the synthetic dataset generated by the [SynLecSlideGen pipeline](https://github.com/synslidegen/synslidegen_pipeline), as presented in the paper [AI-Generated Lecture Slides for Improving Slide Element Detection and Retrieval](https://huggingface.co/papers/2506.23605). It is designed to test the generalization of models trained on synthetic data to real-world lecture slides.

* Project page: [https://synslidegen.github.io/](https://synslidegen.github.io/)
<!-- * Code (generation pipeline): [https://github.com/synslidegen/synslidegen_pipeline](https://github.com/synslidegen/synslidegen_pipeline) -->
<!-- * Dataset repository (on GitHub): [https://github.com/synslidegen/realslide_dataset](https://github.com/synslidegen/realslide_dataset) -->

## How to Download:

### Using Git via terminal
```bash
git lfs install
git clone https://huggingface.co/datasets/nerdyvisky/realslide
```

### Using Python
```python
pip install huggingface_hub
python
from huggingface_hub import snapshot_download

repo_id = "nerdyvisky/realslide"  # your full repo path
local_dir = snapshot_download(repo_id=repo_id, repo_type="dataset")
```

## Overview of RealSlide Benchmark

The RealSlide benchmark consists of 1050 real-world lecture slides collected from Creative-Commons licensed graduate lecture slide decks. Full list [here](https://docs.google.com/spreadsheets/d/1bX05zEv0hyZ-FAvmyTfdMi8pdsPqv2DQGB_AHmIJIzk/edit?usp=sharing) 
Each slide is manually annotated by human-annotators with Slide Object Detection in COCO Format and Text-based slide summary to aid benchmarking VLMs for Slide Image related tasks.


<!-- ![Overview of RealSlide](https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realslide_overview.jpg) -->

## Dataset Components

The dataset includes samples for two main tasks, each with manually verified annotations:

<!-- * **RealDet (Slide Element Detection):** Contains real lecture slides with annotations for various elements within slides (e.g., titles, text, images). -->
<!-- * **RealRet (Text Query-based Slide Retrieval):** Contains real lecture slides suitable for retrieval tasks based on text queries, enabling models to retrieve relevant slides based on textual content. -->

<!-- ### RealDet Samples -->
<!-- <table border="1"> -->
<!-- <tr> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realdet1.png" alt="RealDet1" width="100%"></td> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realdet2.png" alt="RealDet2" width="100%"></td> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realdet3.png" alt="RealDet3" width="100%"></td> -->
<!-- </tr> -->
<!-- </table> -->

<!-- ### RealRet Samples -->
<!-- <table border="1"> -->
<!-- <tr> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realret1.png" alt="RealRet1" width="100%"></td> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realret2.png" alt="RealRet2" width="100%"></td> -->
<!-- <td><img src="https://raw.githubusercontent.com/synslidegen/synslidegen_pipeline/main/code/assets/realret3.png" alt="RealRet3" width="100%"></td> -->
<!-- </tr> -->
<!-- </table> -->

## Usage

This dataset can be used for evaluating models trained on synthetic datasets or for fine-tuning models for lecture slide element detection and retrieval. The data is provided with manually verified annotations, making it suitable for benchmarking and performance evaluation.

## Citation

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

```bibtex
@article{maniyar2025ai,
  title={AI-Generated Lecture Slides for Improving Slide Element Detection and Retrieval},
  author={Maniyar, Suyash and Trivedi, Vishvesh and Mondal, Ajoy and Mishra, Anand and Jawahar, CV},
  journal={arXiv preprint arXiv:2506.23605},
  year={2025}
}
```