We study implementation issues for spatial convolution filters and their Fo
urier alternative, with the aim to optimize the accuracy of filter output.
We focus on Gaussian scale-space filters and show that there exists a trade
-off scale that subdivides the available scale range into two subintervals
of equal length. Below this trade-off scale Fourier filtering yields more a
ccurate results than spatial filtering; above it is the other way around. T
his should be contrasted with demands of computational speed, which show th
e opposite tenet.