The problem of reconstructing a depth map from a sequence of differently focused images (focal stack) is called Depth from focus. The core idea of this method is to analyze the sharpness of each pixel and compare it along the $z$ axis of the focal stack to estimate the true depth value. This approach has two main drawbacks: it depends on the optics of the camera and on the focus measure operator. Recent advances in deep learning techniques show promising results in this way, however, still have problems generalizing to different scenes, cameras, optics and focal stack focus positions. In this paper we propose a novel deep learning based method to approach this problem. Firstly, we propose to estimate distances relative to the focal stack focus position instead of estimating true or invariant depths, allowing us to generalize different scenes and optical setups without losing the possibility to extract real distances. Secondly, we present our novel architecture: a 2D siamese encoder–3D decoder with a differentiable argmax regression that is able to compute depth from stacks of variable sizes. Finally, we compare our method with 2 other depth from focus algorithms and with a monocular depth estimation method.

从一系列不同聚焦的图像（焦点堆栈）重建深度图的问题称为焦点深度。该方法的核心思想是分析每个像素的锐度，并沿$z$焦堆栈的轴来估计真实的深度值。这种方法有两个主要缺点：它取决于相机的光学系统和焦点测量操作员。深度学习技术的最新进展以这种方式显示出有希望的结果，但是，仍然存在推广到不同场景、相机、光学和焦点堆栈焦点位置的问题。在本文中，我们提出了一种新的基于深度学习的方法来解决这个问题。首先，我们建议估计相对于焦点堆栈焦点位置的距离，而不是估计真实或不变的深度，使我们能够概括不同的场景和光学设置，而不会失去提取真实距离的可能性。其次，我们展示了我们的新颖架构：具有可微 argmax 回归的 2D siamese 编码器 - 3D 解码器，能够从可变大小的堆栈计算深度。最后，我们将我们的方法与来自聚焦算法的其他 2 个深度和单目深度估计方法进行了比较。