We propose to estimate the shift gradient based on the corresponding pixel importance using a non-linear map- ping function. Let the importance be normalized such that S(x, y) ∈ [0, 1]. The mapping function is desired to result in bigger gradient when the importance is closer to 0 and smaller value for importance closer to 1. The mapping is non-linear so that the value drops faster when the impor- tance gets closer to 1. An intuitive choice for such a func- tion is the zero-mean Gaussian,