Welcome to the SEE project repository! We introduce a new framework for enhancing images across a broad light range using event-based cameras. Our approach adjusts brightness dynamically to enhance image quality in both low and high light conditions. This repository contains the data and code for the SEE-600K dataset and SEE-Net framework.

In this project, we explore how to use event-based cameras, with their high dynamic range (HDR), to handle images under various lighting conditions. We propose the SEE-600K dataset—a large-scale dataset capturing images under a range of lighting conditions—and a novel framework, SEE-Net, that can adjust brightness smoothly and effectively using events.

In our research paper, we:

• Collect the SEE-600K dataset consisting of 610,126 images and corresponding events under 202 scenarios with a range of lighting conditions.
• Introduce the SEE-Net framework, which effectively adjusts image brightness using event-based data.
• Demonstrate that our framework performs well across a broad range of lighting, from very low to very high light levels.
• Show the flexibility of our method through pixel-level brightness adjustments, which can be useful for various post-processing applications.

The datasets supporting the results of this article are publicly available:

• SEE-600K Dataset:
  • OneDrive
  • Huggingface
• SDE Dataset: SDE Dataset GitHub

The SEE-Net framework is designed to adjust the brightness of images under a variety of lighting conditions. It utilizes event-based data to enhance low-light and high-light images and improve their overall quality.

Key Features:

• Brightness Adjustment 🌈: Our framework allows for pixel-level brightness adjustment based on the events captured by the camera.
• Compact & Efficient ⚡: SEE-Net is designed with efficiency in mind, with only 1.9 million parameters, making it suitable for real-time applications.

To get started with this repository, clone it using the following command:

git clone https://github.com/yunfanLu/SEE.git

Next, install the required dependencies:

pip install -r requirements.txt

You can then use the code to process images using our SEE-Net framework and adjust their brightness using the SEE-600K dataset.

1. Download the SEE-600K dataset

   Download the SEE-600K dataset using the link provided above.

2. Data preprocessing and alignment Before training, the raw data must be properly extracted and aligned.

   • For data extraction, follow the step-by-step instructions in the tools/ directory.
   • For temporal and spatial alignment, please refer to the IMU-based registration tool: https://github.com/yunfanLu/IMU-Registration-Tool This tool provides millisecond-level alignment between RGB frames and events.
   • If you download the dataset from Hugging Face, the data has already been aligned and can be used directly without additional registration.

3. Train the SEE-Net model We provide two training configurations, SEE and SDE, both located in the options/ directory.

   • The full training command is already written in options/SEE/SEENet_SEE.sh.

   • Before running, modify the paths in the script to match your local environment (dataset path, log directory, dependencies).

   • A typical training command is:

     python see/main.py \
       --yaml_file="options/SEE/SEENet_SEE.yaml" \
       --log_dir="./logs/SEE/SEENet_SEE/" \
       --alsologtostderr=True

   • During training, the model learns to reconstruct and adjust image brightness under a wide range of lighting conditions using event guidance.

4. Brightness adjustment with exposure prompts After training, the SEE model supports continuous brightness control via an exposure (brightness) prompt B.

   • By changing the value of B, the model can generate images with different exposure levels from the same input.

   • Internally, this is implemented by sampling different exposure prompts and decoding the corresponding outputs:

     # 3. more exposure prompt
     exposure_prompt = torch.ones(size=(B, 1, 1, 1)) * exposure_B
     exposure_prompt = exposure_prompt.to(images.device)
     exposure_normal_reconstructed = self._decoding(light_inr, exposure_prompt)

   • In evaluation mode, increasing B results in brighter outputs, while decreasing B produces darker images.

   • This design enables pixel-level, continuous brightness adjustment as a post-processing step, without retraining the model.

5. Notes

   • The optimal operating range of the brightness prompt is typically around B ≈ 0.5, which corresponds to normal exposure.
   • Extreme values of B are mainly intended for analysis and visualization, rather than as target outputs.
   • The framework is designed for brightness adjustment rather than full HDR reconstruction.

Here are some example outputs generated using our framework:

• Low-Light Enhancement 🌙: Images processed under very low-light conditions.
• High-Light Recovery 🌞: Images processed under overexposed or bright light conditions.

I released the registration tool separately to this repository - Registration-Tool.

If you use the SEE-600K dataset or SEE-Net in your research, please cite our paper:

@article{lu2025SEE,
  title={SEE: See Everything Every Time - Adaptive Brightness Adjustment for Broad Light Range Images via Events},
  author={Yunfan Lu, Xiaogang Xu, Hao Lu, Yanlin Qian, Pengteng Li, Huizai Yao, Bin Yang, Junyi Li, Qianyi Cai, Weiyu Guo, Hui Xiong},
  year={2025},
}

We welcome contributions to this project! If you have any ideas or improvements, feel free to fork this repository and create a pull request. 💪

• Bug fixes
• New features
• Improvements to documentation

If you have any questions or need assistance, feel free to reach out to us!

• Yunfan Lu: ylu066@connect.hkust-gz.edu.cn
• GitHub: yunfanLu

✨ We hope you find the SEE project helpful and look forward to your contributions! 😄