Validating Autofocus Algorithms with Automated Tests

摘要

For an automated camera focus, a fast and reliable algorithm is key to its success. It should work in a precisely defined way for as many cases as possible. However, there are many parameters which have to be fine-tuned for it to work exactly as intended. Most literature only focuses on the algorithm itself and tests it with simulations or renderings, but not in real settings. Trying to gather this data by manually placing objects in front of the camera is not feasible, as no human can perform one movement repeatedly in the same way, which makes an objective comparison impossible. We therefore used a small industrial robot with a set of over 250 combinations of movement, pattern, and zoom-states to conduct these tests. The benefit of this method was the objectivity of the data and the monitoring of the important thresholds. Our interest laid in the optimization of an existing algorithm, by showing its performance in as many benchmarks as possible. This included standard use cases and worst-case scenarios. To validate our method, we gathered data from a first run, adapted the algorithm, and conducted the tests again. The second run showed improved performance.