Depth estimation and blur removal from a single out-of-focus image

This paper presents a depth estimation method that leverages rich representations learned from cascaded convolutional and fully connected neural networks operating on a patch-pooled set of feature maps. Our method is very fast and it substantially improves depth accuracy over the state-of-the-art alternatives, and from this, we computationally reconstruct an all-focus image and achieve synthetic re-focusing, all from a single image. Our experiments on benchmark datasets such as Make3D and NYU-v2 demonstrate superior performance in comparison to other available depth estimation methods by reducing the root-mean-squared error by 57% & 46%, and blur removal methods by 0.36 dB & 0.72 dB in PSNR, respectively. This improvement is also demonstrated by the superior performance using real defocus images.