Horizons are visible boundaries between certain sediment layers in seismic data, and a fault is a crack of horizons and it is recognized in seismic data by the discontinuities of horizons layers. Interpretation of seismic data is a timeconsuming manual task, which is only partially supported by computer methods. In this paper, we present an automatic method for horizon correlation across faults in 3d seismic data. As automating horizons correlations using only seismic data features is not feasible, we reformulated the correlation task as a non-rigid continuous point matching problem. Seismic features on both sides of the fault are gathered and an optimal match is found based on geological fault displacement model. One side of the fault is the floating image while the other side is the reference image. First, very prominent regions on both sides are automatically extracted and a match between them is found. Sparse fault displacements are then computed for these regions and they are used to calculate parameters for the fault displacement model. A multi-resolution simulated annealing optimization scheme is then used for the continuous point matching. The method was applied to real 3D seismic data, and has shown to produce geologically acceptable horizons correlations.