Performance of feature-based texture classifiers versus Convolutional Neural Network under sparse training data

In this work, we compare the performance of three local-feature-based texture classifiers and a Convolutional Neural Network (CNN) at face recognition with sparse training data. The texture-based classifiers use Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Scale Invariant Feature Transform (SIFT), respectively. The CNN uses six convolutional layers, two pooling layers, two fully connected layers, and outputs a softmax probability distribution over the classes. The dataset contains 100 classes with five samples each, and is partitioned so there is only one training sample per class. Under these conditions, we find that all three feature-based approaches significantly outperform the CNN, with the HOG-based approach showing especially strong performance.