Pairwise Macropixel Comparison Can Work at Least as Well as Advanced Holistic Algorithms for Face Recognition

It is well known that, due to illumination effects and the registration/alignment problem, it does not make sense to compare the “values" of two single-pixels for face recognition. But does that mean that the comparison of two “big" pixels makes no sense either? This paper shows that, by taking a few pixels together as one “big" pixel, called macropixel, and measuring the similarity of macropixels by simple Euclidean distance, a method that counts best matched macropixels indeed works very well for face recognition ‐ experiments show that it is not only much better than traditional holistic algorithms, but is also at least comparable with recently developed ones, if not better. The superiority of the extremely naive macropixel counting approach over wellestablished ones stimulates us to rethink: Does the seemingly dedicated process of our brains for face pattern recognition involve dimensionality reduction? Has the current advance of computer vision research touched the underlying problem in face recognition?