Optimization of a Scale-Invariant Feature detector using scale-space scans

Abstract

High-resolution scale-space scanning is introduced as a feature-probing technique in difference-of-Gaussian detectors. Scans of the feature response are produced versus scale-space parameter σ for different window sizes, for a set of diverse images. Mean repeatability scans are used to select the filter parameters of a reliable Scale-Invariant Feature Transform (SIFT) detector. A simple and hardware-friendly feature descriptor is also proposed and is tested in relation with the proposed optimized detector. This study can guide design optimizations without degradation of the detector response, especially in real-time systems.