The use of prior assumptions in human visual processing.

在人类视觉处理中使用先验假设。

• 批准号: EP/D039916/1
• 负责人: Wendy Adams
• 金额: $ 15.99万
• 依托单位: University of Southampton
• 依托单位国家: 英国
• 项目类别: Research Grant
• 财政年份: 2006
• 资助国家: 英国
• 起止时间: 2006 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=EP%2FD039916%2F1
• 关键词: use; prior; assumptions; human; visual

The complex pattern of light entering our eyes as we view the world is used to recover the three-dimensional structure of our environment. To do this quickly and efficiently our visual system makes certain assumptions. For example, in the absence of other information, we assume that objects are illuminated from above ('light-from-above prior'). Imagine an egg sitting in the sun. Shading makes the egg bright on the top and darker underneath. Without seeing the light-source we assume that it is overhead and use this assumption to help determine the shape of the egg, i.e. convex. In contrast, if you held up an open egg-box, each depression would be dark at the top and brighter at the bottom, helping you to see it as concave. The 'light-from-above' prior helps us to estimate the shape of an object from its shading pattern: a process called 'shape-from-shading' (SFS). We are interested in whether visual assumptions ('priors') are innate or learnt in response to interactions with the environment. Chickens always interpret shading patterns as if the light source was overhead, even if raised in an environment where light always comes from below. In contrast, humans trained in an environment where the light-source is always to one side learn to interpret objects accordingly. A series of experiments will investigate just how adaptable the human visual system is. Observers will view and touch virtual objects via a 'robot arm' that allows us to manipulate whether the objects feel convex or concave. With this method we will train people to believe that objects are lit from the side or from below. We will measure how this training affects the interpretation of other, novel objects, and how long the effects last. We can thus uncover how and over what time-scale the human visual system adapts to changes in its environment.A concave object (e.g. dimple) is easy to find amongst many convex objects (e.g. buttons), based on differences in the shading patterns. This convex / concave visual search task is easiest if the objects are lit from above, consistent with the 'light-from-above' prior. This provides a useful method to investigate the 'light-from-above' prior and how it interacts with other visual scene information.The mechanisms for implementing the 'light-from-above' prior could be very simple or quite complex. For example, evidence is mixed on whether the 'light-from-above' prior moves with the head, such that if you were upside-down you would assume that objects were lit from below. We can investigate this by asking observers to perform visual search tasks with their head or body at different orientations. We can also ask whether SFS is affected by context. For example, if surrounding objects are clearly lit from the side, then the visual system might also interpret central objects as if they are lit from the side. Visual search performance may also change when observers actively move the light source.Visual search behaviour will be compared to shape judgements (convex vs. concave) and judgements about how light or dark objects appear. All of these tasks use the 'light-from-above' prior but may be differentially affected by contextual manipulations. For example, head position might affect shape judgements but not visual search. Differences between the tasks will inform us about the divergence of mechanisms behind various judgements involving shape and shading.This work is of great relevance to scientists interested in the assumptions that the visual system uses; little research has investigated how priors are learnt, how general they are and how they are affected by contextual information. More broadly, anyone using remote presentation of visual or touch information (e.g. virtual reality, remote surgery) should be aware of the assumptions that observers use to interpret displays. For example, display orientation and previous experience could dramatically affect 3D perception, especially under diminished cue conditions.

当我们观察世界时，进入我们眼睛的复杂光线模式被用来恢复我们环境的三维结构。为了快速有效地做到这一点，我们的视觉系统会做出一些假设。例如，在没有其他信息的情况下，我们假设物体是从上面照亮的（“光从上面优先”）。想象一个鸡蛋坐在阳光下。阴影使鸡蛋顶部明亮，底部较暗。在没有看到光源的情况下，我们假设它在头顶上，并使用这个假设来帮助确定鸡蛋的形状，即凸形。相比之下，如果你举起一个打开的蛋盒，每个凹陷在顶部是黑暗的，在底部是明亮的，帮助你看到它是凹的。“上方光线”先验知识帮助我们从物体的阴影模式中估计物体的形状：这一过程称为“阴影形状”（SFS）。我们感兴趣的是视觉假设（“先验”）是先天的还是在与环境的相互作用中习得的。鸡总是解释阴影模式，如果光源是在头顶上，即使在一个环境中，光总是来自下面。相比之下，在光源总是在一侧的环境中训练的人类学会了相应地解释物体。一系列的实验将研究人类视觉系统的适应性。观察者将通过一个“机器人手臂”来观察和触摸虚拟物体，这个手臂允许我们操纵物体是凸还是凹。通过这种方法，我们将训练人们相信物体是从侧面或从下面照亮的。我们将测量这种训练如何影响对其他新奇物体的解释，以及这种影响持续多久。因此，我们可以揭示人类视觉系统如何以及在什么样的时间尺度上适应环境的变化。基于阴影图案的差异，凹对象（例如酒窝）很容易在许多凸对象（例如按钮）中找到。如果物体是从上方被照亮的，那么这种凸/凹视觉搜索任务是最简单的，与“上方的光”先验一致。这提供了一个有用的方法来研究“上面的光”先验以及它如何与其他视觉场景信息相互作用。“上面的光”先验的实现机制可以非常简单，也可以非常复杂。例如，证据是混合的，关于“来自上方的光”是否与头部一起移动，这样，如果你是颠倒的，你会假设物体是从下方照亮的。我们可以通过要求观察者以不同的方向进行头部或身体的视觉搜索任务来调查这一点。我们还可以问SFS是否受上下文的影响。例如，如果周围的物体从侧面被清楚地照亮，那么视觉系统也可能将中心物体解释为从侧面照亮。当观察者主动移动光源时，视觉搜索性能也可能发生变化。视觉搜索行为将与形状判断（凸与凹）以及对亮或暗物体如何出现的判断进行比较。所有这些任务都使用“上方的光”先验，但可能会受到上下文操作的不同影响。例如，头部位置可能会影响形状判断，但不会影响视觉搜索。任务之间的差异将告诉我们的分歧机制背后的各种判断涉及形状和shading.This工作是非常相关的科学家感兴趣的假设，视觉系统使用;很少有研究调查先验知识是如何学习的，如何一般，他们是如何受到影响的上下文信息。更广泛地说，任何使用远程呈现视觉或触摸信息（例如虚拟现实，远程手术）的人都应该知道观察者用来解释显示的假设。例如，显示方向和先前的经验可能会显着影响3D感知，特别是在减弱的提示条件下。

项目成果

Light priors, learning and feedback

简单的先验、学习和反馈

• 发表时间: 2007
• 作者: W Adams

Efficient Visual Recalibration from Either Visual or Haptic Feedback: The Importance of Being Wrong

• DOI: 10.1523/jneurosci.2749-10.2010
• 发表时间: 2010-11-03
• 期刊: JOURNAL OF NEUROSCIENCE
• 影响因子: 5.3
• 作者: Adams, Wendy J.;Kerrigan, Iona S.;Graf, Erich W.
• 通讯作者: Graf, Erich W.

Combining specular and diffuse lighting to recover 3-D shape

结合镜面反射和漫射照明来恢复 3D 形状

• 发表时间: 2007
• 作者: W Adams