Towards Embedded Robot Vision for Multi-scale Object Recognition - Repeatability of Interest Points Detected in Half-octave Binomial Pyramids

The construction of multi-scale image pyramids is used in state-of-the-art methods that perform robust object recognition, such as SIFT and SURF. However, building such image pyramids is computationally expensive, especially when implementations in embedded systems with limited computing resources are considered. Therefore, the use of alternative less expensive approaches are necessary if near real-time operation is desired. Previous work has reported that using binomial filters to construct half-octave multi-scale pyramids consumes only 1/4 of the processing time of the Gaussian pyramid originally used in the SIFT framework. Here we investigate how interest points detected using the binomial approach behave when compared to the Gaussian approach, focusing on repeatability. Experimental results show that in average up to 86% of interest points detected with the original SIFT pyramid building scheme are also detected when using the binomial method, despite of large gains in processing time. When rotation of image features is considered, experimental results demonstrate that slightly superior repeatability of interest points is achieved using the binomial pyramid.