Improving color constancy by selecting suitable set of training images

With very simple implementation, regression-based color constancy (CC) methods have recently obtained very competitive performance by applying a correction matrix to the results of some low level-based CC algorithms. However, most regression-based methods, e.g., Corrected Moment (CM), apply a same correction matrix to all the test images. Considering that the captured image color is usually determined by various factors (e.g., illuminant and surface reflectance), it is obviously not reasonable enough to apply a same correction to different test images without considering the intrinsic difference among images. In this work, we first mathematically analyze the key factors that may influence the performance of regression-based CC, and then we design principled rules to automatically select the suitable training images to learn an optimal correction matrix for each test image. With this strategy, the original regression-based CC (e.g., CM) is clearly improved to obtain more competitive performance on four widely used benchmark datasets. We also show that although this work focuses on improving the regression-based CM method, a noteworthy aspect of the proposed automatic training data selection strategy is its applicability to several representative regression-based approaches for the color constancy problem.