Selecting a background for the training images of a correlation filter: a comparative study

Correlation filters (CF) have been widely used for detecting and recognizing patterns in 2-D images. These filters are designed to yield sharp correlation peaks for desired objects while exhibiting low response to clutter and background. CFs are designed using training images that resemble the object of interest. However it is not clear what should be the background of these training images. Some methods use a white background while others use the mean value of the target region. It is important to determine an appropriate background since a mismatched background may cause the filter to discriminate based on the background rather than the target pattern. In this paper we discuss a method to choose training images, and we compare the effects of different backgrounds on the filter performance in different scenarios using both synthetic (pixels in the background chosen from a Gaussian distribution) and real backgrounds (photographs of different sceneries) for testing. In our comparisons we do not restrict ourselves to using a background with constant pixel intensity for training but also include in the training images backgrounds with varying pixel intensity with mean and standard deviation equal to the mean and standard deviation of the target region. Experiments show that without a prior knowledge of the background in the testing images, training the filters using a background with the mean and variance of all the desired objects tends to give better results.