Shape and texture aware graph processing

Neural network training is time-consuming and often application specific. While pre-trained convolutional neural networks can be fine-tuned for other applications, the lower levels are often affected by changes in rotation and scale, since the approach is purely appearance-based. However, hybrid approaches that take both shape and appearance into account have shown success in general computer vision, so in this work we propose a graph network architecture that captures shape and texture information. Our approach separates shape and texture within a graph neural network in a manner analogous to an active appearance model (AAM), offering potential robustness to shape and scale. To test the effectiveness of this approach, we perform a series of experiments in both unsupervised and supervised settings on the MNIST and CIFAR datasets. First, we perform experiments using unsupervised image data and node position reconstruction; we also perform a cross-database experiment. As part of this work, we introduce an extension to the spectral graph convolution approach GCN by automatically generating weights based on a local and global representation in an effort to make the convolutions more flexible and less application or data specific. We call this new approach Generative GCN (GenGCN). Second, we develop approaches to allow grouping and ungrouping of triangle nodes within the graph. Specifically, we propose a Similarity Matrix generation approach to correctly identify similar triangle nodes. We also introduce a variant of the DIFFPOOL approach that allows us to reverse the grouping process. We performed experiments on MNIST with respect to triangle node grouping and ungrouping. Finally, we perform classification tasks on MNIST and CIFAR, experimenting on a variety of architectural variations and parameters.