Focus Depth Anything: Towards Robust Depth from Focus in the Wild

Abstract

Depth from Focus (DfF) estimates scene depth from multi-focus images. However, continuous depth is observed only through discrete focus responses, yielding an ill-posed inverse problem with multiple plausible solutions. Existing DfF methods mainly improve the accuracy of focus measurements, yet rarely address the discretization-induced ambiguity when mapping local focus responses back to continuous depth, which limits robustness in the wild. Therefore, we introduce Focus Depth Anything (FDA), a Bayesian-inspired posterior-refinement framework. Specifically, FDA treats a pretrained depth estimation foundation model as a strong depth prior and leverages multi-focus consistency as physics-grounded cues to locally correct this prior, thereby suppressing spurious solutions that fit the discrete focus cues equally well. To make the posterior correction sensitive to subtle focus variations, we propose the Semantic–Focus Disentanglement (SFD) module, which disentangles focus and semantic representations and produces multi-focus consistency residuals. Moreover, to reinforce the coupling between the depth prior and focus responses, we develop a physically based Point Spread Function (PSF) simulator that generates large-scale multi-focus images to effectively support model learning. Extensive experiments demonstrate that FDA achieves robust depth estimation in the wild, consistently outperforming existing DfF methods. The code and dataset are available at https://github.com/jiangfeng-Z/Focus-Depth-Anything.