Segmentation and depth perception of naturalistic textures by human observers

在日常生活中，我们的视觉系统每时每刻都在不知不觉中毫不费力地将杂乱的视觉场景图像解析成离散的3D表面和对象。我们的大脑究竟是如何实现这一惊人壮举的，人们对此知之甚少，尽管关于视觉通路早期阶段的神经元如何提取和表示简单的图像特征的知识非常丰富。自然图像通常由纹理丰富的表面组成，其许多小特征极大地增加了这些计算挑战。特别是，我们很难理解纹理特征是如何集成和/或分离到视觉图像的不同区域中的，以及它们的深度关系是如何推断的。 我研究的长期目标是对视觉系统如何编码和分离属于不同表面的图像特征并确定它们之间的深度关系有一个明确的理解。由此产生的知识将是我们对人类视觉系统基本理解的重大进步，并将在未来计算机视觉和人工智能技术的设计中发挥不可估量的价值。 建议的实验将建立在我们在三种人类心理物理任务中使用纹理刺激的最新进展：感知纹理密度的同时对比，相邻纹理之间的分割(图形-背景分离)，以及来自运动视差和立体视觉的深度感知。我们将使用更多的自然纹理刺激，以展现真实世界图像的实际复杂性，并确保结果对真实情况的适用性。这项工作还将涉及基于生物现实组件和操作的计算建模，以预测心理物理反应-从而提供对该模型总结或捕获人类行为的程度的严格、定量测试。 所有拟议的项目都将为研究生和博士后提供高度技术性的培训。为了进行这项研究，他们必须学习和使用关于线性和非线性系统和信号的理论知识(例如空间卷积、线性滤波、傅立叶谱、非线性等)、用于创建和呈现视觉刺激的计算机编程(MatLab)(心理物理学工具箱)，以及在某些情况下用于模拟模型。目标3将涉及机器学习，目标4将涉及高级计算机图形学。

From moment to moment in everyday life, our visual system unconsciously and effortlessly parses images of cluttered visual scenes into discrete 3D surfaces and objects. Exactly how our brains achieve this amazing feat is remarkably poorly understood, in spite of a wealth of knowledge about how neurons in early stages of the visual pathways extract and represent simple image features. Natural images are often comprised of richly textured surfaces, whose many small features greatly compound these computational challenges. In particular, we poorly understand how texture features are integrated and/or segregated into distinct regions of the visual image, and how their depth relationships are inferred. The long-term goal of my research is to achieve a well-defined understanding of how the visual system is able to encode and segregate image features belonging to distinct surfaces and to ascertain their depth relationships. The resultant knowledge will be a major advance in our fundamental understanding of the human visual system, and will be invaluable in the design of future computer vision and artificial intelligence technology. The proposed experiments will build upon our recent progress using texture stimuli in three kinds of human psychophysical task: simultaneous contrast of perceived texture density, segmentation ("figure-ground segregation") between adjacent textures, and depth perception from motion parallax and stereopsis. We will use more naturalistic texture stimuli to bring out the actual complexities of real-world images and to insure the applicability of the results to realistic situations. The work will also involve computational modeling, based on biologically realistic components and operations, to predict psychophysical responses - thus providing a rigorous, quantitative test of how well the model summarizes or captures human performance. All of the proposed projects will provide highly technical training to graduate students and post-doc's. To pursue this research, they will have to learn and use theoretical knowledge about linear and nonlinear systems and signals (e.g. spatial convolution, linear filtering, Fourier spectra, nonlinearities, etc), computer programming (Matlab) both for creating and presenting visual stimuli (Psychophysics Toolbox), and in some cases for simulating models. Aim 3 will involve machine learning, and Aim 4, advanced computer graphics.