CAREER: Articulated Model Recognition

本研究解决了从真实图像中检测和识别可变形的铰接形状（如人和动物）的问题。目的是开发自顶向下和自底向上的识别策略，其中自底向上检测特征，对它们进行分组并访问对象模型，而自顶向下使用对象模型搜索特征。对象以从形状轴模型派生的树状图结构表示。存储在节点中的信息包括概率，用于指定物体和成像函数的可能几何变形。树形图结构允许根据树匹配操作来描述咬合和关节。为了进一步降低任务的计算复杂度，研究了一种信息理论准则来确定图像中物体信息的特征。PI还将开发和教授两门课程，其中一门是纽约大学的新课程，以支持计算机视觉方面的一个重点项目；这些课程将是“视觉导论”和“分组与识别”的专业课。

This research addresses the problems of detection and recognition of deformable articulated shapes, such as people and animals, from real images. The aim is to develop a Top-down and Bottom-up recognition strategy, where the Bottom-up detects features, group them, and access the object model, while the Top-down uses the object model to search for features. Objects are represented in a tree-graph structure derived from a shape axis model. The information stored in the nodes include probabilities, used to specify likely geometric deformations of the object and of the imaging function. The tree-graph structure allows for a description of occlusions and articulations in terms of tree matching operations. To further reduce the computational complexity of the task, an information theory criteria is studied to determine features which convey information about the objects in the image. The PI will also develop and teach two courses, one of them new to NYU, to support a focused program in computer vision; these course will be``Introduction to Vision'', and a specialized course in ``Grouping and Recognition''.