Unified Object and Image Representation in 3D Frequency Domain

摘要IIS-9876904Ben-Arie，美国伊利诺伊大学，76,672-12美元。在3D频域中统一对象和图像表示拟议的研究是针对几个主要的视觉问题开发一种频域对象表示和相关方法。具体地，将使用由PI引入的体积频域对象表示(VFR)，其封装了以观众为中心的表示和以对象为中心的表示。VFR包括作为原始表示的2-D切片的对象视图的连续体。该视图表示基于傅里叶切片定理。以前的频域表示是针对二维图像而开发的，并且没有考虑三维物体结构。用于视觉任务的传统2-D频域或空间域图像表示仅限于2-D解释。另一方面，该方法考虑了完整的三维物体结构，因此可以处理不同的问题，如视点或光照无关的识别，关节三维运动等。VFR不仅包括物体外观的连续体，还包括三维结构信息。因此，可以从单个图像导出3-D对象结构。通过直接索引到复数(实数和虚数)体积特征向量空间，可以使用VFR来实现姿势不变和光照不变的目标识别。提出了两种运动估计方法。结果要么是显式定义运动的3-D姿势的时间序列，要么是3-D运动流场。这种三维方法避免了二维流场的问题，特别是在三维旋转运动的估计中，其中实际的运动向量表现为在图像平面上的投影。还计划将VFR模型扩展到几个视觉应用中，如基于三维结构的图像分割、姿势不变和光照不变的目标和人脸识别、单幅图像的三维表面恢复和关节运动识别。

Abstract IIS-9876904Ben-Arie, JezekielUniversity of Illinois$76,672 - 12 mos.Unified Object and Image Representation in 3D Frequency DomainThe proposed research is to develop a frequency domain object representation and related methodology for several major vision problems. Specifically, a volumetric frequency domain object representation (VFR) introduced by the PI will be used, which encapsulates both viewer centered and object centered representations. The VFR includes a continuum of the views of the objects as 2-D slices of the original representation. This view representation is based on the Fourier Slice Theorem. Previous frequency domain representations were developed for two dimensional images and did not consider the three dimensional object structure. Conventional 2-D frequency domain or spatial domain image representations for vision tasks is limited to 2-D interpretation. On the other hand, the proposed approach considers the full three dimensional object structure and therefore can handle diverse problems such as viewpoint orillumination independent recognition, articulated 3-D motion, etc. The VFR includes not only a continuum of object appearances but also 3-D structural information. Thus, 3-D object structure can be derived from single images. Pose invariant and illumination invariant object recognition can be implemented using the VFR by directly indexing into a complex (real and imaginary) volumetric eigenvector space. Two approaches are proposed for motion estimation. The result would be either a temporal sequence of 3-D poses which explicitly define the motion or a 3-D motion flow field. This 3-D approach avoids the problems of 2-D flow fields, especially in the estimation of 3-D rotational motions where the actual motion vectors exhibit themselves as projections onto the image plane. It is also planned to extend the VFR model towards several vision applications, such as image segmentation based on 3-D structure, pose invariant and illumination invariant object and face recognition, 3-D surface recovery from single image and articulated motion recognition.