Learning in Graphical Models
Learning in Graphical Models
复制标题
DOI：
10.1007/978-0-387-69942-4_4
复制
发表时间：
1999
期刊：
影响因子：
--
通讯作者：
Christopher M. Bishop
中科院分区：
文献类型：
--
作者：
Christopher M. Bishop
多媒体内容分析的一个重要特征是多媒体对象表现出比简单对象更丰富的结构。在某些情况下，我们可能需要标记一组相互关联的实例，因为确定对象的类标签取决于空间，时间相关对象的类标签。例如，词性标注，也称为语法标注，是自动确定文本中每个单词的语法角色（或属性）的过程。这是一个典型的问题，如果不检查单词本身和同一个句子中的相邻单词，就无法解决，因为自然语言中的许多单词在不同的时间可以代表一个以上的词性。另一方面，从棒球视频节目中检测“本垒打”事件是另一个典型的问题，其需要帧序列的联合标记，因为这样的事件由跨越许多视频帧的动作序列组成。如果不检查其视觉内容和序列中的上下文，就无法确定每个帧的标签。
An important characteristic of multimedia content analysis is that multimedia objects exhibit much richer structures than simple objects. In some cases, we might need to label a set of inter-related instances altogether because determining the class label of an object depends on the class labels of spatially, temporally related objects. For example, the part-of-speech tagging, also called grammatical tagging, is the process of automatically determining the grammatical role (or attribute) of each word in a text. This is a typical problem that can not be solved without examining both the word itself and the neighboring words in the same sentence, because many words in natural languages can represent more than one part of speech at different times. On the other hand, detecting ”home run” events from a baseball video program is another typical problem that requires a joint labeling of a sequence of frames, because such a event is composed of a sequence of actions that span over many video frames. The label of each frame can not be determined without examining both its visual content and its context within the sequence.