Probabilistic Models for Image and Video Parsing

计算机视觉中的一个长期存在的问题是能够将图像或视频解析成块，这些块可与场景中的单个对象或至少部分对象相对应。 例如，在1983年，两位计算机视觉研究人员，A. Witkin和J. Tenenbaum解释了人类似乎如何将组织强加于视觉场景（图像或视频），从而押注图像的哪些部分与其他部分相匹配，即使他们不知道他们在看什么。 令人惊讶的是，这种预先强加的，有点盲目的数据组织往往被证明是有用的解释场景的内容。 为了构建越来越灵活和适应性强的计算机视觉系统，我们必须在它们开始理解它们正在看的东西之前，首先对视觉数据进行类似的组织。 以前的相关研究工作包括图像分割，边界概率，轮廓分组，以及在视频数据的情况下，运动分割的算法的发展。 所提出的研究的目标将是显着增加提取的图像块的描述能力，并研究其组装算法。 测试是确定提取的图像块是否形成用于类特定对象识别的有用基础。 这可以在标准基准上测量，该标准基准测量人工视觉系统在典型快照中定位和标记牛、汽车、羊、自行车等的能力。

A long-standing problem in computer vision is to be able to parse an image or a video into chunks which plausibly correspond to individual objects in the scene or, at least, parts of objects. For example, in 1983 two computer vision researchers, A. Witkin and J. Tenenbaum, explained how humans seem to impose an organization on visual scenes (images or video), thereby betting on which image parts go with other parts, even when they have no idea what they are looking at. The surprising thing is that this pre-imposed, somewhat blind, organization of the data often turns out to be useful for interpreting the contents of the scene. In order to build increasingly flexible and adaptable computer vision systems we must design them to first impose a similar organization on the visual data, before they begin to understand what they are looking at. Previous relevant research efforts include the development of algorithms for image segmentation, boundary probability, contour grouping, and, in the case of video data, motion segmentation. The objective of the proposed research will be to significantly increase the descriptive power of the extracted image chunks, and to study algorithms for their assembly. A test is to determine whether the extracted image chunks form a useful basis for class-specific object recognition. This can be measured on standard benchmarks which measure artificial vision systems ability to locate and label cows, cars, sheep, bicycles, and so on, in typical snapshots.