We present the results of relocating 327,000 southern California earthquakes that occurred between 1984 and 2002. We apply time-domain waveform cross-correlation for P and S waves between each event and 100 neighboring events identified from the catalog based on a 3D velocity model. To simplify the computation, we first divide southern California into five polygons, such that there are ∼100,000 events or less in each region. The polygon boundaries are chosen to lie in regions of sparse seismicity. We calculate and save differential times from the peaks in the cross-correlation functions and use a spline interpolation method to achieve a nominal timing precision of 0.001 sec. These differential times, together with existing P- and S-phase picks, are input to the double-difference program of Waldhauser and Ellsworth (2000, 2002) to calculate refined hypocenters. We divide the southern California region into grid cells and successively relocate hypocenters within each grid cell. The overall resulting pattern of seismicity is more focused than the previously determined pattern from 1D or 3D models. The new improved locations are more clustered, in many cases by a factor of two or three, and often show clear linear alignments. In particular, the depth distribution is improved and less affected by layer boundaries in velocity models or other similar artifacts.