Disparity probability volume guided defocus deblurring using dual pixel data

In this paper, we address the problem of defocus deblurring from dual pixel image pair. The defocus blur has the inherent characteristic that the blur amount is related to depth of the scene. The left and right views of dual pixel data exhibits the depth dependent disparity cues. However, previous method using dual pixel data lacks to exploit the disparity cues of the left and right dual pixel images. Therefore, we propose to leverage the disparity information in defocus deblurring network. For this, we propose the disparity probability volume module which predicts the pixel-wise disparity probability in unsupervised manner. We then incorporate the disparity probability volume into defocus deblurring network to utilize the spatially varying blur amount information explicitly. The experiment shows that our disparity probability volume can improve the performance on existing methods on real-world images in terms of both visual quality and image quality metrics.