An Introduction to Graph Neural Networks in Computer Vision: A Practical Perspective

Abstract

While deep learning models such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViT) have achieved remarkable success in computer vision, they are inherently limited to processing data with fixed and regular structures, such as 2D grids. However, Graph Neural Networks (GNNs) extend the convolutions to non-Euclidean domains, allowing them to process data structured as graphs. This capability enables GNNs to exploit complex relationships beyond the fixed spatial organization of pixels. Another important aspect is that GNNs are particularly well-suited for semi-supervised learning scenarios, especially those with few labeled data points. This is a significant advantage in computer vision, where obtaining large volumes of labeled data for training deep learning models is usually necessary. This paper provides a practical overview of GNNs and their application to visual data, with a focus on image classification, retrieval, segmentation, and detection. We begin with foundational concepts and representative GNN models. Building on this foundation, we examine how visual graphs can be constructed from deep features and how GNNs can be applied to different computer vision tasks. Finally, we highlight current trends, key challenges, and promising directions for future research on GNNs in computer vision.