Limitations and next steps
As graph deep learning continues to make strides in CV, several challenges and promising research directions have begun to emerge. One of the primary challenges in applying graph-based methods to CV is scalability.
Scalability issues in large-scale visual datasets
As we saw in Chapter 5, as the size of visual datasets continues to grow, constructing and processing large graphs becomes computationally expensive. For instance, a high-resolution image represented as a pixel-level graph could contain millions of nodes, making it challenging to perform graph convolutions efficiently.
Researchers are exploring various approaches to address this issue. One promising direction is the development of more efficient graph convolution operations. For example, the GraphSAGE algorithm can be used with a sampling-based approach to reduce the computational complexity of graph convolutions. Another approach is to use hierarchical graph representations, where the...
