Everyone loves Local Binary Patterns! They're so simple and effective. But their basic parameters were pretty much hand-picked by their creator; I wondered if it was possible to learn a better Local Binary Pattern-like descriptor using supervised data. Since this was before everyone went crazy for Deep Learning, I came up with a scheme using tree-structured or fern-structured quantizers that had vanilla Local Binary Patterns as a special case. The basic insight is that a tree (or list, in the case of ferns) of binary decisions can be seen as mapping to a feature space consisting of binary strings.