Question

求翻译，急要（英语可能会有打错的）

The discriminative approaches,such as Conditional Random Fields, have been successfully used in structural labeling tasks.In the previous chapters,we addressed the representation issue of context in the discriminative framework.One main limitation of discriminative methods is that they require a detailed-labeled data set to construct the models.When labeling information is unreliable,vague or missing,it is hard to apply them.It is nearly impossible to find a dataset where this is not the case in real-world vision applications.Labeling every pixel of an image manually is very tedious，and not practical for real-world datasets that include large numbers of images.

On the other hand,it is easier to obtain weakly-labeled image data,such as images with captions.While different types of weakly labeled datasets are available for image labeling,we focus here on image data with multiple levels of labels.In many cases,the label values have different levels of granularity,and they can be grouped into a label hierarchy based on their semantics.For example,a region with the label’hippo’can also be labeled as ‘animal’ or ‘animate object’.Using such more abstract or coarse labels requires less effort in labeling images,as the coarser label set often has a simpler structure than the detailed ones,and is easier to specify.Figure 5.1 shows a typical example where the coarse label configuration has simpler boundaries.We also consider another aspect of the coarseness in labeling,which means some images regions are unlabeled.Those regions are either not relevant to target problem,or too vague to be recognized.In practice,it is desirable to augment the existing detailed-labeled images with those coarsely-labeled image to leverage the learning process.

Answer 1

歧视性的方法，如条件随机域，已成功地应用于结构标签的任务。在前面的章节中，我们讨论的上下文的表示问题的歧视性的框架。的判别方法的一个主要限制是，他们需要一个详细的标记的数据集构建模型。当标签信息是不可靠的，模糊的或思念，很难将它们。它找到一个数据集，这是没有的情况下，真实世界的视觉的应用几乎是不可能的。标记图像的每一个像素的大小是非常烦琐而又真实世界的数据集，包括大量的图像不实际的。
另一方面，它是更容易获得弱标记的图像数据，如图像和字幕。而不同类型的弱标记的数据集的图像标记，我们这里重点对标签的多层次的图像数据。在许多情况下，标签值有不同的粒度级别，并且它们可以分在标签的层次结构的基础上的语义。例如，一个与label'hippo'can地区也被称为“动物”或“动画”的对象。使用这种更抽象的或粗糙的标签需要较少的努力，在标记的图像，如粗糙的标签集通常有一个较详细的结构更简单，更容易指定图5.1显示了一个典型的例子。在粗标签配置简单的边界。我们也考虑在标记的粗糙的另一个方面，这意味着一些图像区域是未标记的。这些地区是目标问题不相关的，或过于模糊难以辨认。在实践中，它是理想的增强现有的详细的标签这些粗糙的标记图像的ED利用学习过程。

望采纳，谢谢