A Multi-Scale Generalization of the HoG and HMAX Image Descriptors for Object Detection

Recently, several powerful image features have been proposed which can be described as spatial histograms of oriented energy. For instance, the HoG [5], HMAX C1 [14], SIFT [13], and Shape Context feature [4] all represent an input image using with a discrete set of bins which accumulate evidence for oriented structures over a spatial region and a range of orientations. In this work, we generalize these techniques to allow for a foveated input image, rather than a rectilinear raster. It will be shown that improved object detection accuracy can be achieved via inputting a spectrum of image measurements, from sharp, fine-scale image sampling within a small spatial region within the target to coarse-scale sampling of a wide field of view around the target. Several alternative feature generation algorithms are proposed and tested which suitably make use of foveated image inputs. In the experiments we show that features generated from the foveated input format produce detectors of greater accuracy, as measured for four object types from commonly available data-sets. Finally, a flexible algorithm for generating features is described and tested which is independent of input topology and uses ICA to learn appropriate filters.