Predicting the Attractiveness of Real-Estate Images by Pairwise Comparison using Deep Learning

Abstract

Even images of the same object, people's impression can vary greatly with the image capture method and angle. In this study, we focused on the exterior images of real-estate to predict which of two images of the same property is more attractive. In general, using a model that regresses the attractiveness score of the property image is common; however, it is difficult to stably evaluate a property image attractiveness using an exact attractiveness value. Therefore, instead of using exact values to evaluate the more attractive image, generating a model correctly predicting which is the more attractive one from the two property images of the same property is better. We created a dataset and developed a model that ensures similar output values for image pairs with similar attractiveness and different output values for pairs with different attractiveness. The proposed method predicted the more attractive image with higher accuracy than conventional models.