Feature Group Utility in Model-Based 3D Object Recognition Systems

计算机视觉中最苛刻的应用之一是基于模型的3D物体识别。这种系统通常使用对应技术来识别图像中的物体；场景特征组（例如，点、曲线、曲面或体）和模型特征兼容组之间形成绑定集。单个特征（如点、曲线、曲面或体积）和兼容的模型特征组的区分能力（或效用）。这些系统中单个特征组的识别能力（或效用）是计算机视觉研究中一个具有挑战性和重要的问题。当识别系统期望从数百或数千个数据库中选择正确的模型时，如果要有效地识别对象，索引方法（快速拒绝系统进一步考虑的模型数据库的子集）是有用的，或者通常是必不可少的。使用特性组为执行此索引任务提供了一种很有前途的技术。本研究探讨了特征组效用概念的两个方面：从输入图像中提取的特征组的显著性，以及数据库中对象模型固有对称性所产生的特征的对称排列。并对初步实验结果进行了描述。所提出的研究的具体目标是：(a)将显著性和对称性概念置于坚实的理论基础上，(b)将现有的识别显著性和对称特征组的技术推广到当前（多面体或二次曲面）的领域，使其成为更具表现力和一般性的特征类。(c)扩展目前超过100个对象模型的数据库，以包括额外的，更真实的对象。

One of the most demanding applications in computer vision is model-based 3D object recognition. Such systems often use correspondence techniques to identify objects in an image; sets of bindings are formed between groups of scene features (e.g., points, curves, surfaces, or volumes) and compatible groups of model features. The discrimination ability ( or utility) of individual features (e.g. points, curves, surfaces, or volumes) and compatible groups of model features. The discrimination ability (or utility) of individual feature groups in these systems is a challenging and important problem in computer vision research. When recognition systems are expected to select the correct model(s) out of a database of hundreds or thousands, indexing methods (which quickly reject a subset of the model database from further consideration by the system) are useful or often essential if objects are to be recognized efficiently, and the use of feature groups offers a promising technique for performing this indexing task. This research investigates two facets of the feature group utility concept: saliency of feature groups extracted from the input image, and symmetric arrangements of features arising from inherent symmetries of the object models in the database. Preliminary experimental results are also described. Specific goals of the proposed research are to (a) place both the saliency and symmetry concepts on a firm theoretical basis, (b) generalize existing techniques for identifying salient and symmetric feature groups out of their current domain (of polyhedral or quadric surfaces into more expressive and generic classes of features, and (c) expand a current database of over 100 object models to include additional, more realistic objects.