Depth From Focus via Test-Time Optimization of Monocular Depth Estimation Models

Depth from focus (DFF) estimates scene depth by analyzing images captured at different focus distances. Recent deep learning–based DFF methods can predict depth at a metric scale; however, their accuracy is often limited by the relatively small amount of available training data. To overcome this limitation, we propose a more accurate DFF framework that leverages prior knowledge from monocular depth estimation (MDE) models trained on large-scale datasets. Specifically, at test time, the output of a existing DFF method is used as a reference, and the parameters of the MDE model are optimized on a per-scene basis. Experiments on synthetic and real-world datasets demonstrate that the proposed method improves both depth accuracy and structural quality, and achieves consistent improvement across a wide range of scenes compared to existing DFF approaches.