Jasinschi R S
Computer Vision Laboratory, University of Maryland, College Park 20742.
Biol Cybern. 1991;65(6):515-23. doi: 10.1007/BF00204665.
Energy filters are tuned to space-time frequency orientations. In order to compute velocity it is necessary to use a collection of filters, each tuned to a different space-time frequency. Here we analyze, in a probabilistic framework, the properties of the motion uncertainty. Its lower bound, which can be explicitly computed through the Cramér-Rao inequality, will have different values depending on the filter parameters. We show for the Gabor filter that, in order to minimize the motion uncertainty, the spatial and temporal filter sizes cannot be arbitrarily chosen; they are only allowed to vary over a limited range of values such that the temporal filter bandwidth is larger than the spatial bandwidth. This property is shared by motion sensitive cells in the primary visual cortex of the cat, which are known to be direction selective and are tuned to space-time frequency orientations. We conjecture that these cells have larger temporal bandwidth relative to their spatial bandwidth because they compute velocity with maximum efficiency, that is, with a minimum motion uncertainty.
能量滤波器被调谐到时空频率方向。为了计算速度，有必要使用一组滤波器，每个滤波器都调谐到不同的时空频率。在这里，我们在概率框架下分析运动不确定性的特性。其下限可以通过克拉美 - 罗不等式明确计算得出，它将根据滤波器参数具有不同的值。我们针对伽柏滤波器表明，为了使运动不确定性最小化，空间和时间滤波器的大小不能任意选择；它们只允许在有限的值范围内变化，使得时间滤波器带宽大于空间带宽。猫的初级视觉皮层中的运动敏感细胞也具有这一特性，这些细胞已知具有方向选择性并且被调谐到时空频率方向。我们推测这些细胞相对于其空间带宽具有更大的时间带宽，因为它们以最高效率计算速度。也就是说，具有最小的运动不确定性。
