Convolutional neural networks and vision transformers have dominated image classification, but graph neural networks have stayed on the sidelines. A new arXiv preprint, CoViG, argues that GNNs are widely used elsewhere yet remain mostly unexplored in computer vision, and introduces a coarsened vision GNN for image classification.

Because the abstract is truncated, the source does not include details of the architecture or experimental results. The title and opening suggest the work is part of a recent push to apply graph-based reasoning to vision tasks, rather than a full comparison with established models.

With only one source, there are no conflicting accounts to reconcile. The main takeaway is that CoViG positions GNNs as a promising but underused tool for image classification, alongside the better-known CNN and transformer paradigms.