Beyond histograms: why learned structure-preserving descriptors outperform HOG

Statistical image descriptors based on histograms (e.g. SIFT [1], HOG [2]) are widely used in image processing, because they are fast and simple methods with high classification performance. However, they discard the local spatial topology and thus lose discriminative information contained in the image. We discuss the relations between HOG and VNMF descriptors, i.e. structure free histograms versus learned structure-preserving patterns. VNMF is a shift-invariant, sparse, nonnegative unsupervised learning algorithm [8, 9, 5], that provides a distinct decomposition of the input into its parts. The VNMF descriptor outperforms the statistical HOG descriptor, because it preserves spatial topology leading to better classification results on real-world human action recognition benchmarks [11, 12].