Detection of Face Features using Adapted Triplet Loss with Biased data

The task of classification of imbalanced data has recently become a major issue in pattern recognition and machine learning. The fundamental challenge with this type of data is that the smaller classes tend to be more valuable. However, standard classifiers have a bias towards the large classes and disregard the small ones. This results in a poor performance, especially in the minority class where accuracy is low. To demonstrate this problem, we used the VGGFace2 dataset, which is not biased. Therefore, we intentionally biased the data by distributing the images unequally within each class and address the problem of representation learning using triplet loss. We propose a model to obtain an informative feature embedding using the ResNet-18 network, and then use these learned embeddings for image classification. In addition, we use a method to improve the naive triplet loss named adapted triplet, to eliminate the bias resulting from the triplet selection process and to demonstrate the generalization on biased data. We implement this approach using the PyTorch framework. The experimental results show that the proposed approach achieves an accuracy of 90.01% and 96.71% for the triplet and adaptive triplet loss, respectively.