Data-adaptive polarization filtering is used to improve the detection of converted seismic phases. Both direct waves and mode-converted PS and SP arrivals may be more easily picked on the filtered records. An autopicking routine is applied that cuts the polarization filtered traces according to the modelled traveltime of each phase through an initial structure. Use of forward-modelled, source-receiver times reduces the likelihood of an automatic pick being incorrectly made on spurious spikes in the polarization filtered trace. It is therefore a realistic way of automatically picking multiphase data sets or, more generally, linearly polarized phases where low signal-to-noise ratios may be encountered. The method is suitable for any three-component seismic data and is here applied to local earthquakes recorded in North Island, New Zealand. Intermediate energy is observed between the direct P and S arrivals due to phase conversion at the interface between the Indo-Australian and subducting Pacific plates. The amplitudes of these converted arrivals are often too low for them to be identified above the P-wave coda but polarization filtering of the records enables the yield of converted phase picks to be greatly increased.