Early representation of 3D volumetric shape in visual object processing

项目摘要 这个项目的目标是测试一个新的理论框架，以了解腹侧通路如何 有助于观察物体。在标准框架中，对2D图像数据的一系列神经操作通过 许多中间皮层阶段，包括V4区，导致高水平的知觉表征，包括 在腹侧通路的最后阶段，物体身份的表征。然而，我们的初步 来自固定猴的微电极数据显示，V4中的许多神经元代表体积（体积- 包围）3D形状，而不是2D图像图案。这些神经元对许多不同的2D图像做出反应， 相同的3D形状具有不同的形状深度提示，包括阴影，反射和折射。他们甚至 优先响应于传达3D体积形状而没有2D提示的随机点立体图 不管怎样此外，我们的初步结果与双光子功能成像在麻醉猴V4 示出了3D形状信号通过它们相似性被分组，且还与同构（相同轮廓）2D 形状信号（其可以为相应的3D形状推断提供证据）。我们建议 利用这些初步数据，通过展示V4区3D形状调整的普遍性，分析 这些神经元使用的3D形状编码策略，并测量2D和3D形状信号是如何 排列在V4表面的微观水平上。我们期望这些结果能提供有力的证据 3D形状碎片提取是V4处理的主要目标。这种早期的3D形状提取 信息，只有两个突触以外的初级视觉皮层，将建议一个竞争的框架， 理解腹侧通路，其中最初的目标是代表3D物理结构，独立 在这个框架中，物体识别将基于 关于3D物理结构的先前信息，这将解释为什么人类物体识别如此重要。 对图像变化的鲁棒性，在某种程度上，最好的计算视觉系统不是。科学影响 这项工作将使视觉实验转向理解真实的世界三维结构的表示 （而不是2D平面刺激），并鼓励计算视觉科学家将早期的3D形状 处理成当前技术水平的深度卷积网络模型。

Project Summary The goal of this project is to test a novel theoretical framework for understanding how the ventral pathway subserves object vision. In the standard framework, a series of neural operations on 2D image data through many intermediate cortical stages, including area V4, leads to high-level perceptual representations, including representation of object identity, at the final stages of the ventral pathway. However, our preliminary microelectrode data from a fixating monkey show that many neurons in V4 represent volumetric (volume- enclosing) 3D shape, not 2D image patterns. These neurons respond to many different 2D images that convey the same 3D shape with different shape-in-depth cues, including shading, reflection, and refraction. They even respond preferentially to random dot stereograms that convey 3D volumetric shape with no 2D cues whatsoever. Moreover, our preliminary results with 2-photon functional imaging in anesthetized monkey V4 show that 3D shape signals are grouped by their similarity, and also group with isomorphic (same outline) 2D shape signals (which could contribute evidence to corresponding 3D shape inferences). We propose to capitalize on these preliminary data by demonstrating the prevalence of 3D shape tuning in area V4, analyzing the 3D shape coding strategies used by these neurons, and measuring how 2D and 3D shape signals are arranged at a microscopic level across the surface of V4. We expect these results to provide strong evidence that extraction of 3D shape fragments is a primary goal of V4 processing. This early extraction of 3D shape information, just two synapses beyond primary visual cortex, would suggest a competing framework for understanding the ventral pathway, in which the initial goal is to represent 3D physical structure, independent of the various 2D image cues used to infer it. In this framework, object recognition would be based on preceding information about 3D physical structure, which would explain why human object recognition is so robust to image changes, in a way that the best computational vision systems are not. The scientific impact of this work would be to divert vision experiments toward understanding representation of real world 3D structure (rather than 2D planar stimuli) and to encourage computational vision scientists to incorporate early 3D shape processing into the deep convolutional network models that are the current state of the art.