Scale Invariance for Evolved Interest Operators

This work presents scale invariant region detectors that apply evolved operators to extract an interest measure. We evaluate operators using their repeatability rate, and have experimentally identified a plateau of local optima within a space of possible interest operators Ω. The space Ω contains operators constructed with Gaussian derivatives and standard arithmetic operations. From this set of local extrema, we have chosen two operators, obtained by searching within Ω using Genetic Programming, that are optimized for high repeatability and global separability when imaging conditions are modified by a known transformation. Then, by embedding the operators into the linear scale space generated with a Gaussian kernel we can characterize scale invariant features by detecting extrema within the scale space response of each operator. Our scale invariant region detectors exhibit a high performance when compared with state-of-the-art techniques on standard tests.