Performance measurement of a general multi-scale template matching method

Generally, template matching methods are very sensible to image orientation, rotation or the size of the template. However, there are known occasions, when they perform better then other keypoint-based techniques. This article demonstrates a massively parallelizable method to multi-scaled template matching. By implementing it sequentially on CPU and on GPU with a naive approach, the runtimes can be measured during a test with multiple scale sizes. With the use of the GPU, the runtimes are 16-times lower. These results indicate that there is place for further research, and to develop a GPU associated solution to handle real-time video processing.