Intelligent Vision Algorithm Selection

From HIT Lab NZ

Researchers:	Adrian Clark
Date:	December, 2005 - Ongoing
Funding:

---

Contents 1 Publications 2 Additional Content 2.1 Image Gallery 2.2 Videos

There are many computer vision algorithms, each specifically tailored to problems within a certain operating environment. The aim of this project is to build a system which can determine the properties of the environment in which it is operating, and then decide which vision algorithm is most appropriate to use.

Each operating environment can be decomposed into a number of "metrics", which define characteristics about the scene, such as hue variance, connected components, etc. A neural network is trained to recognise which vision algorithms are appropriate for dealing with certain combinations of these metrics. At run time, the metrics are extracted from the scene, and the neural network then estimates which vision algorithms have the highest probability of success.

Computer vision research tends to focus on very specific problem domains. Collecting several solutions for limited domains together allows for a larger operating range than any single approach, while still providing the accuracy of a finely tuned algorithm. Using a modular system design, algorithms can be added, removed and updated as advances are made in the science.

As the system is comprised of a library of existing algorithms, it will be as accurate as the algorithms it implements while operating in the domain the algorithms were designed for. However, when outside of a specific domain, the system will be able to estimate which algorithm would perform most accurately. This will result in a system which is overall more accurate than any of it's component algorithms.

Any application that requires the use of computer vision in a range of environments, or environments where conditions are not tightly controlled (such as public spaces), could benefit from this system. The project could also be ideal for people who wish to develop applications utilising computer vision without having to study vision algorithms, as it gives them a robust system which can automatically decide which algorithms would will best meet their needs.

In future more algorithms, metrics and filters can be added to the system to further increase decision making accuracy. Ideally the system would eventually be as knowledgable as a computer vision guru, and able to make well informed decisions about it's choice of algorithms.

Publications

Additional Content

Image Gallery

Videos

There are currently no videos uploaded...