A Hybrid Strategy for Illuminant Estimation Targeting Hard Images

Illumination estimation is a well-studied topic in computer vision. Early work reported performance on benchmark datasets using simple statistical aggregates such as mean or median error. Recently, it has become accepted to report a wider range of statistics, e.g. top 25%, mean, and bottom 25% performance. While these additional statistics are more informative, their relationship across different methods is unclear. In this paper, we analyse the results of a number of methods to see if there exists 'hard' images that are challenging for multiple methods. Our findings indicate that there are certain images that are difficult for fast statistical-based methods, but that can be handled with more complex learning-based approaches at a significant cost in time-complexity. This has led us to design a hybrid method that first classifies an image as 'hard' or 'easy' and then uses the slower method when needed, thus providing a balance between time-complexity and performance. In addition, we have identified dataset images that almost no method is able to process. We argue, however, that these images have problems with how the ground truth is established and recommend their removal from future performance evaluation.