Rule-of-Thirds Detection with Interpretable Geometric Features

Abstract

The Rule-of-Thirds (RoT) is a fundamental compositional principle in visual arts that is a geometric clue for aesthetic questions. While it can be defined formally, the application shows high variance between raters. Existing computational approaches for RoT detection classify images using classifiers trained on heuristic saliency methods or deep learning models. Neither approach explicitly addresses the relation between semantic and structural content, which is necessary to address the subjective component of the problem. We propose a hybrid approach that bridges this gap. Our method leverages a ResNet50-based saliency to bootstrap pre-trained object detection and segmentation networks. We use these outputs to construct interpretable geometric features that quantify alignment with the RoT grid. This approach achieves competitive performance for binary classification trained on explainable models while providing explicit compositional features that scholars can understand. The model's interpretability enables further analysis of aesthetic problems.