RGB-Event ISP: MATLAB Implementation

Overview of ISP

This repository contains the MATLAB implementation of the RGB-Event ISP pipeline, which processes RAW images into color-corrected RGB images. The codebase is developed as part of the RGB-Event ISP Dataset and Benchmark, which introduces the first event-RAW paired dataset for event-based image signal processing (ISP). The dataset and associated methods were presented in our ICLR 2025 paper.

PyTorch Implementation

For the ‌Benchmark‌ implementation in ‌PyTorch‌, please refer to the dedicated repository github.com/yunfanLu/RGB-Event-ISP-Benchmark.

Features

• RAW Image Processing: Converts APS RAW images into RGB images using a predefined ISP pipeline.
• Controllable ISP Pipeline: Implements black level correction, demosaicing, white balancing, denoising, and color correction.
• MATLAB-Based Processing: Provides a structured and modular ISP framework for MATLAB users.
• Dataset and Evaluation: Supports the new RGB-Event ISP dataset with diverse scenes, lighting conditions, and event-RAW alignment.

Note: The ISP process described in our paper uses a ColorChecker as a prior for color correction and does not incorporate event data for RAW to RGB conversion.

Repository Structure

├── BM3D-master                          # BM3D denoising implementation
├── color-checker-extraction-master       # Color checker extraction module
├── color-checker-location-master         # Color checker localization module
├── color-correction-toolbox-master       # Color correction utilities
├── Quad_Bayer_CFA_Modified_...          # Demosaicing module
├── npy-matlab-master                     # NPY file support for MATLAB
├── extract_error_in_log                  # Log analysis scripts
├── testdata-rawdata                      # Sample test data (RAW images and events)
├── apply_cmatrix.m                        # Apply color matrix transformation
├── cfa_pattern.m                          # CFA pattern handling
├── DEMO.m                                 # Main demonstration script
├── DEMO_EVENT.m                           # Demonstration with event integration
├── DEMO_EVENT_LABELME.m                   # LabelMe-based demonstration
├── DngToJpg.m                             # Convert RAW DNG images to JPG
├── DngToPng.m                             # Convert RAW DNG images to PNG
├── get_color_card_coords_from_json.m      # Extract color checker coordinates from JSON
├── Load_Data_and_Metadata_from_DNG.m      # Load RAW data and metadata from DNG files
├── Main.m                                 # Main processing pipeline
├── README.md                              # Documentation (this file)
├── RGBE_ISP.m                             # Core MATLAB function for ISP
├── TASK_VIDEO_FOLDER_RGBE_ISP.m           # Batch processing script
├── TASK_VIDEO_FOLDER_RGBE_ISP_LIMITED_BBOX.m  # Processing script with bounding box constraints
├── tone_curve.mat                         # Predefined tone curve for color correction
├── wbmask.m                               # White balance mask
├── demo.sh                                # Example batch job script for HPC

Installation and Setup

1. Install MATLAB (R2022a or later) with Image Processing and Signal Processing toolboxes.
2. Clone the repository:

   git clone https://github.com/yunfanLu/RGB-Event-ISP.git
   cd RGB-Event-ISP

3. Add the repository to your MATLAB path:

   addpath(genpath('RGB-Event-ISP'));

4. Ensure the necessary dependencies (BM3D, color correction toolbox) are included in the MATLAB path.

Processing Frames

To process an entire folder of RAW data:

TASK_VIDEO_FOLDER_RGBE_ISP('dataset/Event-Video-Color-Dataset', 20, 0.45, 1, false);

Citation

If you use this code or dataset in your research, please cite our ICLR 2025 paper:

@inproceedings{lu2025rgb,
  title={RGB-Event ISP: The Dataset and Benchmark},
  author={Yunfan Lu and Yanlin Qian and Ziyang Rao and Junren Xiao and Liming Chen and Hui Xiong},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025}
}

License

This project is released under the Academic Research License. It is strictly for non-commercial, research purposes only. Any commercial usage, redistribution, or modification for profit-oriented activities is strictly prohibited.

We welcome researchers and students in the event-version community to use our dataset. Please send me an email using the school email and state the purpose of non-commercial use. I will reply to you within 24 hours!

Contact

For any questions or inquiries, please contact Yunfan Lu at [email protected].