Handwritten digit recognition models are
normally tested with one benchmarking data set like MNIST,
where classical pipeline models are achieving saturated levels
of accuracies; this in-domain test however cannot give a good
insight about their out-of-domain generalization. In this paper,
we investigate cross-dataset generalization from MNIST to
EMNIST-Digits by utilizing a classical pipeline containing
Histogram of Oriented Gradients (HOG) feature extraction,
feature space augmentation, Principal Component Analysis
(PCA) and Support Vector Machine (SVM) classifier. Our
proposed method is Variance Gap Guided Feature
Augmentation (VGFA). This method scales feature level
augmentation noise using the HOG variance of the source
domain along with normalized variance gap between source
and target domain. For the full-dataset testing, our SVM model
without any augmentation achieves 88.955% EMNIST
accuracy with 9.685 percentage point accuracy loss from
MNIST dataset, while VGFA (β=0.7, λ=0.1) achieves 90.3875%
EMNIST accuracy with 8.3225 percentage point loss from

MNIST.