Analysis of interest point distribution in SURF octaves

Speeded-Up Robust Features (SURF) is a state-ofthe- art, scale- and rotation-invariant feature extraction technique with the potential for real-time execution. Although SURF has been extensively employed for multi-scale computer vision applications since its inception, there are still some areas of this computationally complex algorithm that have not been fully explored and require detailed analysis to enable algorithm-level optimization of SURF for real-time execution. In particular, the distribution of interest points in SURF octaves is a topic that requires thorough investigation. Contrary to the present perception, this paper demonstrates that there is a possibility of higher octaves being more significant than the lower octaves in terms of detected interest points for real-life images. The paper also shows that variation of blob response threshold has a significant effect on interest point distribution. The results presented highlight the need of developing a systematic approach to SURF octave selection.