SGER: Depth from Physics in Computer Vision

这个奖项，在探索性研究小额拨款模式下，将检验和测试一种第一原理方法，该方法仅使用关于介质和在单一视图中获得的强度的知识来确定图像中物体的相对深度。计算机视觉中推断深度(或范围)的大多数方法依赖于使用多个摄像头，并且通常需要识别结果图像中的对应点。或者，测距仪的方法依赖于焦距的清晰度测量，这需要校准的焦距镜头和对焦的明确定义。计划中的方法是基于光在半透明漫射介质中的传输和衰减的物理学。这些研究人员已经证明，在理论上和一组小实验中(在模拟生物组织中)，当结构嵌入这种介质中时，可以直接推断结构的相对深度。这项工作特别具有探索性，因为最初的结果是作为其他研究的副产品出现的，因此它在更一般情况下的适用性程度尚不清楚。要探讨的两个问题是：模型的局限性(例如，最小可检测深度差)和敏感性(对假设和参数变化)是什么；物体的大小和形状对确定其相对深度有什么影响。这项研究的目的是开发一种能力，以明确地识别图像中结构的相对深度，从而防止混淆哪些结构覆盖其他结构。因为一次查看就足够了，所以只要介质的性质已知，就有可能进行追溯鉴定。这在评估图像时应该很有用，例如，在水下、生物和大气应用中，以及在更混浊的环境中。

This award, in the Small Grants for Exploratory Research mode, will examine and test a first-principles approach for determining the relative depths of objects in images using only knowledge about the medium and the intensities acquired in a single view. Most methods for inferring depth (or range) in computer vision have depended on the use of multiple cameras and normally require the identification of corresponding points in the resulting images. Alternatively, rangefinder methods depend on sharpness-of-focus measures that require calibrated-focus lenses and an unambiguous definition of focus. The planned approach is based on the physics of light transmission and attenuation in a translucent diffuse medium. These investigators have shown, in theory and with a small set of experiments (in simulated biological tissue), that it is straightforward to infer the relative depths of structures when they are embedded in such a medium. This work is particularly exploratory because the original results arose as a byproduct of other research and so the extent of its applicability in a more general setting is not known. Two of the questions to be probed are: what are the limitations (e.g., minimum detectable depth difference) and sensitivities (to assumptions, and to changes in parameters) of the model; and what is the effect of the size and shape of the object on determination of its relative depth. The research is aimed at the development of a capability to identify unambiguously the relative depths of structures in an image, and thus to prevent confusion about which structures overlie which others. Because a single view is sufficient, it will be possible to make this identification retrospectively, as long as the properties of the medium are known. This should be useful in evaluating images arising, for example, in underwater, biological, and atmospheric applications, as well as in more-turbid environments.