Refocus

Post-captured Dynamic Focus & Aperture Control

This repo contains code to run Refocus - an image editing technique to change focus point and f-stop setting in a post-processing manner.

Run Refocus

To reproduce our work, please do the following steps:

1. Clone the repo

   git clone https://github.com/KhanhNgoDuy/Refocus.git
   

2. Prepare your python environment
3. Install necessary packages

   pip install -r requirements.txt
   

4. Run

   python main.py --img_path <path/to/your/test_img>
   

   If the path is not specified, the code defaults to test on our prepared example.

Current results & UI

Methodology

Our solution implements two distinct approaches to address the refocusing task, a physics-based model and a deep learning-based model.

Physics-Based Approach

The physics-based approach leverages the canonical principle in photography: Circle of Confusion.

Deep Learning-Based Approach

Our deep learning-based approach leverages Diffusion Models to learn complex real lens optical effects, which are difficult to model explicitly through physical equations.

Training
Below is our training pipeline, training data is available at NTIRE 2023 Bokeh Effect Transformation.

Testing

Our test data is available here.