use a convolution kernel size based on the stddev rather than a