Describe the feature and motivation

In the norm $E(x) = \sqrt{x^2+\epsilon^2}$ used Brox2004 and implemented in cv::VariationalRefinement Class, the parameter $\epsilon$ is fixed and only used to make the norm differentiable in contrast with the TVL1 one.
This parameter is not just a numerical artifact and can be used to pass from L2 norm to L1 norm continuously as it is often the case in the robust norm (or metrics).
It would be very easy to expose this parameter and thus help us to obtain more consistent results in Optical Flow regarding the noise rejection and the regularization.
Could you expose this parameter as it doesn't change anything by default ?

Additional context

No response