MT_Box

Convolutional Neural Networks (CNN) have achieved a huge success in computer vision tasks. In spite of the fact that some CNN models can out-perform human in many specific tasks, the traditional CNN models can only handle images with fixed dimensions, i.e. width and height. Current state-of-the-art CNN models overcome the issue by resizing input images to maintain the consistency of dimensions among inputs. This approach has the main drawback of over up-sampling or over down-sampling with images whose dimensions are far too different to the standard of the models. Moreover, most of CNN models are built over-parameterized, which means most of them are heavier than necessary. In this paper, we aim to make use of graph neural networks to broaden the very new research field of applying the networks on visual tasks. We propose a position-aware dynamic graph propagation scheme to handle super-pixel images created by popular super-pixel segmentation algorithms. Although our model did not out-perform state-of-the-art traditional CNN models on all datasets due to the inevitable information loss of segmentation step, we achieved a huge improvement in accuracy compared to existing graph neural network methods and achieved state-of-the-art accuracy on Superpixel MNIST-75 dataset by 23% error rate drop; we also reduced the number of parameters by 99% compared to VGG16 model.