7,059 high-quality images captured to support dual-pixel (DP) view synthesis and single-image defocus deblurring tasks. The dataset enables a multi-task learning framework that predicts two sub-aperture views from a single blurry input to improve deblurring accuracy.
Use Cases
- Train a multi-task network to predict sub-aperture views and deblurred images from a single blurry input
- Develop reflection removal algorithms using synthesized high-quality DP views
- Benchmark defocus deblurring performance using the PSNR metric against state-of-the-art baselines
Strengths
- 7,059 high-quality images captured for training DP-view synthesis models
- Includes ground truth for two sub-aperture views to support multi-task learning frameworks
- Supports DP-view synthesis from a single input image with a benchmark of ~39dB PSNR