Minimally Supervised Acquisition of 3D Recognition Models from Large Visual Corpora

摘要IIS-9977206纳尔逊，罗切斯特大学，99,133-12美元。从大型视觉公司获得最低限度的监督3D识别模型这是第一次获得为期三年的持续拨款。这项研究致力于从图像中训练3D对象识别系统，该系统将允许在不构建3D几何模型的情况下识别对象，以及在当前几何识别技术不适用的情况下识别对象。在各种形式中，基于图像的方法是目前最成功的通用方法。两个很少被探讨的问题是：(1)这样的系统是否可以从未标记的图像中训练，以及(2)它们是否可以从不可分割的图像中训练。这些问题是本研究方案的重点。使用PI开发的3D识别系统，以及可能的其他系统，该项目将调查训练、更新和提炼3D表示以从非结构化或半结构化图像中识别，并通过工作原型来演示技术。与培训相关的问题包括基本算法设计、控制表示的增长、处理杂乱和纠正错误。与训练语料库相关的问题包括量化显著图像的所需密度、允许的杂乱和所需的种子信息。与识别系统相关的问题包括处理杂乱、空间要求以及处理分层表示或灵活对象等新功能。将为这一进程的各个阶段制定正式模式。

Abstract IIS-9977206Nelson, RandalUniversity of Rochester$99,133 - 12 mos.Minimally Supervised Acquisition of 3D Recognition Models from Large Visual CorporaThis is the first year award of a three-year continuing grant. This research concerns with training 3D object recognition systems from imagery, which will allow the recognition of objects without constructing a 3D geometric model, and the recognition of objects where current geometric recognition technology does not apply. In various forms, image-based methods are currently the most successful general approach. Two issues that have been little explored are: (1) whether such systems can be trained from imagery that is unlabeled, and (2) whether they can be trained from imagery that is not trivially segmentable. These questions are the focus of this research proposal. Using a 3D recognition system the PI has developed, and possibly others, the project will investigate training, updating, and refining 3D representations for recognition from unstructured or semi-structured imagery and demonstrate the technology through working prototypes.Issues connected with training include basic algorithm design, controlling growth of the representation dealing with clutter, and correcting mistakes. Issues associated with the training corpus include quantifying the required density of salient images, allowable clutter, and required seed information. Issues connected to the recognition system include handling clutter, space requirements, and new capabilities such as dealing with hierarchical representations or flexible objects. Formal models for the various stages of the process will be developed.