In each time window, the dip giving the maximum semblance value is picked for use in the statics estimation and interpolation.
In 3-D, I typically perform the dip estimation on a grid that is coarse compared to the receiver geometry. The number of dips makes this step relatively expensive. A grid fine enough to insure that each trace contributes to a few local slant stacks is usually sufficient.
It is also possible to use beam stacks Biondi (1992) instead of local slant stacks. Beam stacks offer improved resolution as the number of neighboring traces used is increased. However, beam stacks require an extra piece of information, effectively the distance from the source to the stack location. For array studies with unknown source locations and multiple sources present, this information is not available. A procedure that scans over beam stack trajectories would be an interesting alternative, though more expensive that the local slant stack technique because of extra parameters corresponding to the source location.