Ablation study of the approaches for kinship verification

Abstract

The Kinship Verification aims to determine blood relativeness between people from visual data. This problem gains more attention from the research community during the last few years due to low human and machine performance on this task. Multiple attempts were taken to improve the performance of the Kinship Verification, but most of them are done under diverse experimental settings, and the results are benchmarked on different datasets. In our work, we conduct an ablation study to compare these improvements and determine whether different approaches are compatible with each other or should be applied separately. This ablation study consists in testing different feature extractors, embedding combinations, neural network hyperparameters and such data preprocessing techniques as data augmentation, face alignment and image normalization. This thesis showcases the importance of reach feature representation and efficient embedding combination for the overall accuracy of Kinship Verification.