"Local Rank Differences" Image Feature Implemented on GPU

A currently popular trend in object detection and pattern recognitionis usage of statistical classifiers, namely AdaBoost and itsmodifications. The speed performance of these classifiers largelydepends on the low level image features they are using: both on theamount of information the feature provides and the executional time ofits evaluation. Local Rank Differences is an image feature that isalternative to commonly used haar wavelets. It is suitable forimplementation in programmable (FPGA) or specialized (ASIC) hardware,but --- as this paper shows --- it performs very well on graphicshardware (GPU) as well. The paper discusses the LRD features and theirproperties, describes an experimental implementation of LRD in graphicshardware, presents its empirical performance measures compared toalternative approaches and suggests several notes on practical usage ofLRD and proposes directions for future work.