SURF-ing a Model of Spatiotemporal Saliency

Zhai and Shah (2006) proposed a model of spatiotemporal saliency using a combination of temporal and spatial attention models. The temporal model utilized Lowe’s SIFT (2004) to compute feature points and the correspondences between them in successive frames. Bay, Tuytelaars, & Van Gool introduced SURF (2006) as an alternative feature detector and descriptor. The authors of SURF show that it is faster than and superior to SIFT. This investigation replicated the model of Zhai and Shah evaluating performance with SURF used in place of SIFT. The performance of SIFT and SURF temporal model variants was tested on a variety of frame sets. In addition to qualitative comparisons, the SIFT and SURF model variants were quantitatively compared on the speed of both computing feature correspondences, and the resulting speed of homography computation. The robustness of the models to increased temporal spacing between frames was also tested at various step intervals. The SURF model was much faster as expected. Otherwise the models were generally indistinguishable. Both showed little change in performance due to an increase in the time between successive frames.