New synthetic data generation method enables fast, accurate training of gesture recognizers used in rapid prototyping of user interfaces
UCF researchers have developed a method that quickly generates realistic synthetic data and enables gesture recognizers to significantly improve their accuracy. The new method, called Stochastic Resampling (SR), is computationally efficient, has minimal coding overhead, and does not require expert knowledge to implement. SR-generated synthetic samples also outperform those of competitive, state-of-the-art methods, namely Perlin Noise and Sigma-Lognormal Model. In some cases, reducing mean recognition errors by more than 70 percent.
The new method produces synthetic data using 2D original input that is scanned from a hardcopy, drawn with a stylus on an interactive display or made with a finger on a touch device. SR can also be extended to 3D to support full-body, human-computer interaction (gestures and action recognition).
SR is an ideal rapid prototyping tool for user interfaces because it can run in real time, enabling users to generate synthetic data on demand, and it is easily ported to almost any target platform (hardware, operating system or programming language). Additionally, SR can be integrated into existing software with little effort.
SR intelligently selects random points along a 2D trajectory that scales the spaces between the points to create realistic variations of a given sample. For example, given a hand-drawn circle with a time series of K points, SR resamples the series into a fixed number of N points along the series' path. The path distance between points is non-uniform, and the direction vector between each consecutive pair of points is extracted and normalized to unit length. Next, the normalized in-between point direction vectors are concatenated together to create a new set of N points, with the origin of the first vector being at the center of the coordinate system. Thereafter, the resulting series can be translated, scaled, skewed, rotated, and so forth, as necessary.
- Low coding overhead and accurately generates synthetic data, even with little user input
- Can be ported to any target platform such as hardware, operating system or programming language
- Fast, versatile, language independent, and can be integrated with existing recognizers
- Can synthesize low quality samples collected on low-end devices
- Train, test and validate recognizers for gestures, handwriting, shapes, equations and schematics for 2D interfaces (used on touch screens and smartphones, for example)
- Can be extended to 3D to support full-body human-computer interaction