It is very necessary for any design to be measured against standards and benchmarks related to that design. Unfortunately, such a thing has been nonexistent for vision chips so far. While, any analog or digital design is reported with several standard tests and benchmarks, none of the vision chips reported include test results in a standard format. Each vision chip has at most had a few specific test diagrams indicating the functionality of the chip under certain conditions (these conditions are often untold or vaguely described). If vision chips are going to be used in industrial applications, and if they are to be introduced to customers, they should be represented by some standard specifications.
The problem of quantifying the characteristics of a vision chip is very challenging. First of all computer vision algorithms, on which vision chips are based, do not have any unanimously agreed standard for testing. Determining the quality of an algorithm has remained subjective. This has been directly transmitted to vision chips. Determining the quality of the final output of vision chips is also subjective in this sense.
However, as the accuracy of image processing performed on digital computers exceeds that of vision chips, and as algorithms implemented in vision chips are relatively simple, it is quite reasonable to compare the simulation results obtained on a computer to those obtained from the vision chip itself.
The second problem in testing vision chips is due to the diversity of vision tasks. There are many different vision tasks that may be implemented in a vision chip. Defining test patterns and standards to cover all aspects of the functionality of a vision chip, and for each vision task is very difficult.
Despite these problems, there are tests that can be performed on vision chips. These tests can easily quantify the reliability, dynamic range for different operations, speed, and so on. Here I introduce a few tests which can be performed on any vision chip. Before these, another important issue, i.e. test conditions are addressed.
In any case, when reporting the functionality of any vision chip quoting absolute distance metrics would be meaningless, unless the exact characteristics of the optical system are given, as optical systems can vary from microscopes to telescopes!. All parameters which should include distance should be reported using ``pixel unit''. For example, for motion detection chips a sensitivity of say 0.1 foot/s is meaningless, while a value of 0.1 pixel/s clearly indicates the performance of the chip, and the system designer can adopt this value to any absolute metric velocity sensitivity with the use of proper optics.