The work that has been published in the specific area of recognition of individual signs is still limited. Murakami and Taguchi considered ten signs using a recurrent neural net and obtained accuracy of 96 per cent ([MT91]). Charayaphan and Marble considered 31 ASL symbols, but only sampled each sign once and simulated the variation and consistently got 27 out of the 31 correct, with the remaining four sometimes correct using a Hough transform. Starner considered 40 signs used in brief sentences and ([Sta95,SP95]) obtained accuracies of 91.3 per cent on raw signs and 99.2 per cent by using a very strict grammar for sentences. More detailed information can be found in [Kad95].
Two datasets of sign language data were captured. The first (original) data was captured for [Kad95], using very cheap equipment. The second, collected for this work, used much higher quality equipment. We discuss the differences between the two sources. For convenience the older dataset will be termed the Nintendo data, and the newer dataset will be termed the Flock data.