EAGER: A New Framework for Balancing Deformability and Discriminability in Computer Vision

在形状匹配、目标识别等计算机视觉问题中，变形性和可识别性往往是两个相互冲突的因素。例如，已经观察到强变形不变描述符对于类别识别的区分能力通常较低。这个急切的项目探索了一种新的框架，用于平衡计算机视觉任务的变形性和区分性。该框架将对象均匀地嵌入到称为方面空间的高维空间中，该对象可以是2D形状、点集、图像、3D体积或表面。然后使用嵌入参数来控制变形不敏感度。从理论和应用两个方面对该框架进行了研究。在该框架的基础上，本项目旨在开发另外三个研究目标：通过自适应地选择变形能力来实现稳健的形状匹配方法，通过在框架中处理关节来实现稳健的点集配准方法，以及通过在嵌入的方面空间中提取特征来实现稳健的图像匹配。这些目标计划在实际应用中进行评估，包括基于轮廓的树叶数据检索、基于计算机的物理治疗中的3D标记匹配以及基于图像的疾病筛查。该项目旨在解决处理变形和提高可识别性这两个主要问题，这两个问题涉及计算机视觉内外的许多子领域。跨学科应用预计将对包括生物多样性研究、生物医学研究等在内的各个领域产生重大贡献。研究结果，包括代码和数据，通过项目网站公开。

Deformability and discriminability are often two "conflicting" factors in computer vision problems such as shape matching and object recognition. For example, it has been observed that strong deformation invariant descriptors often suffer from low discriminative powers for category recognition. This EAGER project explores a new framework for balancing deformability and discriminability for computer vision tasks. The framework uniformly embeds an object, which can be a 2D shape, a point set, an image, a 3D volume or a surface, in a high dimensional space named aspect space. The embedding parameter is then used to control the degree of deformation insensitivity. Both the theoretic and application sides of the proposed framework are investigated. Based on the framework, the project aims to develop three additional research goals: robust shape matching methods by selecting deformability adaptively, robust point set registration methods by dealing with articulation in the framework, and robust image matching by extracting features in the embedded aspect space. These goals are planned to be evaluated on real applications including silhouette-based foliage data retrieval, 3D marker matching in computer-based physical therapy, and image-based disease screening. The project aims to bridge the two main problems, handling deformation and improving discriminability, which relate to many subfields inside and outside computer vision. The interdisciplinary applications are expected to generate significant contributions to various fields including biodiversity studies, biomedical study, etc. The research results, including code and data, are made public available through the project website.