Salient Point Detector (2010)


The core spatiotemporal salient point detector is based on the method developed by Oikonomopoulos et al. (SMC-B’2006). In this updated version of the detector, a set of HOG/HOF descriptors is also extracted around each detected salient point. Each HOG/HOF descriptor is a 1x48 vector (4 blocks along the spatial dimension, 1 block along the temporal dimension, 12 bins per each block histogram), giving rise to a 102-dimensional descriptor around each detected point (also counting the 6-dimensional vector containing information over the spatiotemporal location of the detected point (3d), its spatiotemporal scale (2d) and its saliency value (1d)).