One-Shot-Learning Gesture Recognition using HOG-HOF Features

摘要

The purpose of this paper is to describe one-shot-learninggesture recognition systems developed on the ChaLearn GestureDataset (ChaLearn). We use RGB and depth images and combineappearance (Histograms of Oriented Gradients) and motiondescriptors (Histogram of Optical Flow) for parallel temporalsegmentation and recognition. The Quadratic-Chi distance familyis used to measure differences between histograms to capturecross-bin relationships. We also propose a new algorithm fortrimming videos---to remove all the unimportant frames fromvideos. We present two methods that use a combination of HOG-HOFdescriptors together with variants of a Dynamic Time Warpingtechnique. Both methods outperform other published methods andhelp narrow the gap between human performance and algorithms onthis task. The code is publicly available in the MLOSSrepository. color="gray">