Description

The source of the data is the raw measurements from a Nintendo PowerGlove. It was interfaced through a PowerGlove Serial Interface to a Silicon Graphics 4D/35G workstation. This glove definitely falls into the category of "cheap and nasty". Position information is calculated on the basis of ultrasound emissions from emitters the glove to a 3-microphone "L-Bar" that sits atop a monitor. There are two emitters on the glove; and three receivers. This allows the calculation of 4 pieces of information: x (left/right), y (up/down), z (backward/forward), and roll (is the palm pointing up or down?). x, y and z are measured with 8 bit accuracy. "x, y, z" should not be taken to be the normal 3-dimensional orthogonal basis. In particular, 1 unit in the z direction is not of similar distance to 1 unit in the x or y directions. These x, y, z positions are relative to a calibration point which is when the palm is resting on the seated signer's thigh. Roll is 4 bits. The data is susceptible to occasional "spikes" caused by random ultrasound noise. Median filters have been found to be beneficial in solving this problem. Finger bend is generated by conductive bend sensors on the first four fingers. Values vary between 0 (straight) and 3 (fully bent). Accuracy is 2 bits. The gloves automatically apply a hysteresis filter on these bend sensors. At best, these measurements should be treated sceptically. See past usage for a more detailed discussion on the data collection methodology. The data was collected from five signers: Signer -- Description -- Sessions -- Total samples/sign Adam -- Sign linguist - PhD completed in area. -- 2 -- 8 Andrew -- Natural signer - signing since youth -- 3 -- 8 John -- Professional Auslan interpreter -- 5 -- 18 Stephen -- Professional Auslan interpreter -- 4 -- 16 Waleed -- The researcher. Novice signer -- 4 -- 20 Each session was taken at a different time, after a break, etc. The "adam" dataset were sampled in a fixed order -- this means that they are subject to fatigue effects, etc. All other datasets were sampled in random order. The "waleed" and "stephen" datasets contain signs that begin with "cal-". These were considered as a means of calibration, but didn't work out too well. The data presented is the raw data with no filtering. Occasional dropouts in x, y, z values. These can be easily fixed using a median filter. Average number of frames per instance is 51, but varies from 30 to 102. The data is in a comma separated file containing all of the attributes mentioned above. Each sign sample is stored in a single file. The directory hierarchy is as follows: -Each signer is in a separate directory. -Each session from signer is in a subdirectory. Each session is denoted by a number. -Each sample is in a file named by the sample appended with the number of the sample of that sign. The filenames indicate the class.

Related Papers

  • Mohammed Waleed Kadous and Claude Sammut. The University of New South Wales School of Computer Science and Engineering Temporal Classification: Extending the Classification Paradigm to Multivariate Time Series. [link]
  • [link]

Related datasets