For the FREAK detector in the BoW image classification sample, it is mentioned that using Binary-Split with normalized FREAK features would possibly work better than the k-modes. However, it is not clear in the documentation how this option can actually be implemented/incorporated in the code.