Image Pair Prediction Using GAN-Based Style Encoders: Evaluation with Cosine Similarity and AU-ROC

In order to evaluate the model, the style encoders of GAN, GAN-CLS, GAN-INT, and GAN-INT-CLS were used to predict image pairs in the experiment, with cosine similarity scores used to validate the results, and AU-ROC (averaged over 5-fold cross-validation) reported. As a baseline, the cosine similarity between text features from the text encoder was also calculated in the study.