Depth from Defocus via Direct Optimization

A team of researchers including Holly Jackson and Benjamin Recht has published a significant computer vision paper titled 'Depth from Defocus via Direct Optimization' on arXiv. The work tackles the classic problem of estimating a scene's depth map from a collection of defocused images, a task crucial for computational photography, robotics, and AR/VR.

The core innovation is a global optimization approach made feasible with modern computing. The method is based on alternating minimization. First, with the depth map fixed, the forward blur model becomes linear, allowing the all-in-focus image to be recovered via convex optimization. Second, with the all-in-focus image fixed, the depth at each pixel can be computed independently in an 'embarrassingly parallel' grid search. This alternating cycle between convex optimization and parallel search proves highly effective.

The results challenge the prevailing paradigm. The authors demonstrate that their direct, physics-based optimization method can solve the depth-from-defocus problem at higher resolutions than current state-of-the-art deep learning approaches. They validated the technique on both synthetic and real-world benchmark datasets, showing promising performance compared to prior methods. This suggests that for certain well-defined inverse problems, traditional optimization with sufficient compute can rival or surpass learned models, offering a potentially more interpretable and physically-grounded alternative. The public release of the code allows for immediate verification and application by the community.