Does this mean that MachineBox's algorithm is not good? The short answer is no: we cannot say that the MachineBox algorithm is not good. The longer answer requires a more nuanced understanding that combines engineering understanding and an understanding of machine learning. As far as the algorithm of facebox goes, there are no exact details about what facebox is composed of. But we can deduce what goes on.
First, note that the images with matches are all over 50% in their confidence. We can then assume that facebox considers a match being found only if the confidence level greater than 50%. I verified this by running the recognizer on a directory of over 1,000 images of faces. Only those that are matched have a greater-than 50% confidence. The program is as follows:
func testFacebox() error {
files, err := filepath.Glob("OtherFaces/*"...