EAGER: Grouping Features for Object Localization and Image Search

互联网上的大量图像需要高效的图像搜索算法来帮助用户找到包含感兴趣对象的图像。当前最先进的图像搜索方法通常将感兴趣的对象表示为一组特征，并通过搜索覆盖所需特征的矩形窗口来定位对象。由于不考虑特征之间的空间关系，这些方法的判别能力低，假阳性率高。相反，这个EAGER项目将目标定位描述为一个全局特征分组问题，其中检测到的特征根据一些一般的空间关系进行分组，例如群凸性、边界和内部特征的分离以及特征亲和力。利用新的图模型和方法实现了最优特征分组。在这个公式中，搜索窗口是一个更紧密的边界多边形，而不是矩形。通过考虑空间关系和更紧密的边界多边形，本项目开发的特征分组方法有望比现有的图像搜索方法产生显著的改进，这可以通过在标准数据集上的测试来验证。在这个项目中开发的源代码计划在这个项目完成后向公众开放。本研究的重点是目标定位，这也可以促进许多其他计算机视觉应用，如场景匹配和重建、目标检测和识别、基于内容的视频检索和视频监控。

Numerous images on the internet call for efficient and effective image search algorithms to help users to find images that contain the object of interest. The current state-of-the-art image-search methods usually represent an object of interest as a set of features and localize the object by searching for a rectangular window that covers the desirable features. Without considering spatial relations among the features, these methods usually suffer from a low discriminative power and a high false positive rate. Instead, this EAGER project formulates object localization as a global feature grouping problem, where detected features are grouped according to some general spatial relations, such as group convexity, the separation of boundary and internal features, and feature affinities. The optimal feature grouping is achieved by using new graph models and approaches. Within this formulation, the search window is a tighter bounding polygon rather than a rectangle.By considering spatial relations and tighter bounding polygons, the feature-grouping approach developed in this project is expected to produce a significant improvement over existing image-search methods, which can be verified by testing on a standard data set. The source code developed in this project is planned to be made publicly accessible upon completion of this project. This research is focused on object localization, which can also benefit many other computer-vision applications, such as scene matching and reconstruction, object detection and recognition, content-based video retrieval, and video surveillance.