Abstract: Local features used for the sake of object recognition must be both robust to common image deformations and informative for the purpose of model identification. A recent proposal for such local features, consisting of phase and amplitude responses of complex-valued steerable filters, proved to be robust to common brightness changes, 2-D rotation, and small scale variations. Here, we extend that approach to achieve robustness to large scale changes. We provide empirical results comparing our feature to: a)local features based on differential invariants in \cite{schmid_pami_97}, and b) Scale Invariant Feature Transform (SIFT) local features in \cite{lowe_iccv_99}. The results show that our local features lead to better performance when dealing with common illumination changes, 2-D rotation, sub-pixel translation, and large scale changes than the differential invariants. For illumination changes our features show better performance than SIFT features with similar performances for all other deformations studied. In addition, our feature has fewer dimensions than the SIFT feature (less expensive in terms of database search), and the spatial support is smaller (more robust to skew). Finally, we demonstrate the usefulness of our feature in a simple object recognition system.