Hand posture recognition using compositional techniques

This work proposes a compositional approach to hand posture recognition, using sparse features. The hand posture is decomposed into relevant compositions which are learned for each hand posture class without supervision; no hand segmentations or localization during training is needed. To learn relevant composition prototypes, an entropy range maximization loop was introduced, by performing k-means clustering several times. Experimental results compare favorably with results of both image categorization and hand posture recognition reported in literature.