Median filters are nonlinear rank-order filters based on replacing each element of the source vector with the median value, taken over the fixed neighborhood (mask) of the processed element. These filters are extensively used in image and signal processing applications. Median filtering removes impulsive noise, while keeping the signal blurring to the minimum. Typically mask size (or window width) is set to odd value which ensures simple function implementation and low output signal bias. You can use an even mask size in function calls as well, but internally it will be changed to odd by subtracting 1.
Another specific feature of the median filter implementation in Intel IPP is that elements outside the source vector, which are needed to determine the median value for “border” elements,
are located in a delay line. If the delay line is absent, then they are set to be equal to the corresponding edge element of the source vector.