Waleed Kadous' Sign Language Dataset

Introduction

This data comes from my honours thesis. It consists of sample of Auslan (Australian Sign Language) signs. Examples of 95 signs were collected from five signers. A total of 6 650 sign samples were collected. The data in particular is described in sections 4.3, 4.4. A list of the signs, why they were selected, etc can be found here.

Notes on the data

Signers

The data was collected from five signers:

Signer	Description	Sessions	Samples per signs
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.

Special notes

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.

Files

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.

Data source

The data was collected from a single right-handed Nintendo Power Glove. The following data are sampled at a rate of approximately 23 Hz:

Position

Position is given for x (left/right), y (up/down), z (backward/forward) information. This information is extremely noisy, because the source of the information is an ultrasonic transmitter on the glove. Delays between beeps on the ultrasound transmitter and three receivers arranged in an "L-bar" on the CRT are measured. This means it is susceptible to noise from sound reflections, overhead fluorescent lights and so on. The accuracy on position is 8 bits.

Roll

Roll measures how horizontal the wrist is. The starting point is "palm down", which has a value of 0. Values increase clockwise, to 5 at "palm up" and so on. The maximum value is 11 which occurs when the hand is rotated approximately 30 degrees counterclockwise. Accuracy is slightly less than 4 bits.

Finger Bend

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.

File Format

The file format consists of comma-separated values for the following:

x position between -1 and 1. Units are approximately metres.
y position between -1 and 1. Units are approximately metres.
z position between -1 and 1. Units are not metres. This space should not really be treated as linear, although it is safe to treat it as monotonically increasing.
roll with 0 meaning "palm down", rotating clcokwise through to a maximum of 1 (not included), which is also "palm down".
pitch: has a value of -1, indicating that it is not available for this data. Should be ignored.
yaw: has a value of -1, indicating that it is not available for this data. Should be ignored.
thumb bend: has a value of 0 (straight) to 1 (fully bent).
fore bend: has a value of 0 (straight) to 1 (fully bent).
index bend: has a value of 0 (straight) to 1 (fully bent).
ring bend: has a value of 0 (straight) to 1 (fully bent).
little bend: has a value of 0 (straight) to 1 (fully bent). In this case, it is a copy of ring bend. Should be ignored.
keycode: Indicates which key was pressed on the glove. Should be ignored.
glove state 1 Should be ignored.
glove state 2 should be ignored.
Receiver values: determines if all receivers received values from all transmitters. A value of 0x3F indicates all receivers received information from all transmitters. Other values indicate this is not the case.

The library C source for glovedata.h and glovedata.c may be useful for you in reading/writing the data into a higher level C structure.

Conditions on data use

The data may be used provided that:

The source of the data (http://www.cse.unsw.edu.au/~waleed/tml/data/) is acknowledged.
If you do publish any work using the dataset, I'd like to hear about it.

The data

Ok, finally here's the data. Almost forgot to include it. The data is tarred and gzipped. Its uncompressed size is approximately 58 megabytes. Its compressed size is approximately 2.8 megabytes.

allsigns.tar.gz.

Mohammed Waleed Kadous

Last modified: Sat Aug 15 12:05:07 EST