Single-Image Depth From Defocus With Coded Aperture and Diffusion Posterior Sampling

Abstract

We propose a single-snapshot coded-aperture depth-from-defocus (DFD) method that reconstructs RGBD by posterior-guided diffusion, combining a learned RGBD diffusion prior with a differentiable DFD forward model. Given a measurement, we approximately sample from the posterior induced by the forward model (data fidelity) and the diffusion prior, using a denoise–correct–renoise procedure that applies data-consistency updates in the denoised domain. Unlike U-Net–style regressors, our approach requires no paired defocus–RGBD training data and does not couple training to a specific camera configuration. Experiments on comprehensive simulations and a prototype camera show stable RGBD reconstructions across noise levels, outperforming U-Net baselines and a classical coded-aperture DFD method.Our implementation and experimental configuration will be released publicly upon publication.

Publication
IEEE Transactions on Computational Imaging
Hajime Nagahara
Hajime Nagahara
Professor

He is working on computer vision and pattern recognition. His main research interests lie in image/video recognition and understanding, as well as applications of natural language processing techniques.