An integrated probabilistic framework for shape and surface interpretation

描述（由申请人提供）：形状和表面解释的综合概率框架Jacob Feldman和Manish Singh，Rutgers大学视觉形状的表示是感知的中心问题之一，影响物体识别和场景理解的许多方面。但是，关于人脑如何计算形状表示的全面而有原则的解释还不存在。一个关键的困难是理解大脑如何将图像划分为不同深度的不同表面，解释每个表面的3D形状，并将每个形状划分为不同的部分。在以前的工作中，PI已经开发了新的，原则性的数学方法来理解人类的形状表示，基于贝叶斯估计的形状骨架的想法。形状骨架是形状的轴向结构的表示，与经典的中轴表示相关但不同。中轴表示将形状分解为它们的分量轴，形状关于分量轴近似局部对称。PI的方法将其改写为概率推理问题，与大多数当代神经计算模型一致，但与大多数其他形状计算模型不同。这使得该方法的范围扩大，涵盖了更广泛的问题，图像分解成不同的表面和3D形状的解释。该方法在理论上是统一的，并且在数学上适合于在并行计算架构中实现，使得其作为神经形状编码的模型是合理的。本研究计划的目的是充分发展形状表示作为一个概率估计问题，并测试从这个框架中产生的许多经验预测。具体目标包括扩展和测试的概率方法，形状表示，并将其推广到关键的相关问题，包括数字/地面解释和三维形状。这种扩展将使我们有可能在一个连贯的数学框架中，将形状表征与对视觉图像中表面的理解完全结合起来，这个框架与视觉皮层中的计算是一致的。

DESCRIPTION (provided by applicant): An integrated probabilistic framework for shape and surface interpretation Jacob Feldman and Manish Singh, Rutgers University The representation of visual shape is one of the central problems of perception, influencing many aspects of object recognition and scene understanding. But a comprehensive and principled account of how the human brain computes shape representationsdoes not yet exists. A key difficulty is in understanding how the brain divides the image into distinct surfaces at distinct depths, interprets the 3D shape of each surface, and divides each shape into distinct parts. In previous work the PIs have developed novel, principled mathematical methods for understanding human shape representation, based around the idea of Bayesian estimation of the shape skeleton. The shape skeleton is a representation of the axial structure of the shape, related to though different from classical medial axis representations. Medial axis representations break shapes down into their component axes, about which the shape is approximately locally symmetric. The PI's approach recasts this as a probabilistic inference problem, consistent with most contemporary neurocompuational modeling, but unlike most other computational models of shape. This allows the approach to be expanded in scope, encompassing the broader problem of the decomposition of the image into distinct surfaces and the interpretation of 3D shape. The approach is theoretically unified, and is mathematically suitable to be implemented in a parallel computational architecture, making it plausible as a model of neural shape coding. The aim of this research program is to fully develop shape representation as a probabilistic es- timation problem, and test the many empirical predictions that emanate from this framework. Specific aims include expanding and testing the probabilistic approach to shape representation, and generalizing it to critical related problems, including figure/ground interpretation and 3D shape. This expansion will make it possible to fully integrate shape representation with the comprehension of surfaces in the visual image, in a coherent mathematical framework consistent with what is known about computation in visual cortex.