DiscoveringDynamicSalientRegionsfor Spatio-TemporalGraphNeuralNetworks

In this paper, we propose a novel method to enhance vision Graph Neural Networks (GNNs) by an additional capability, missing from any other previous works. That is, to have nodes that are constructed for spatial reasoning and can adapt to the current input. Prior works are limited to having either nodes attached to semantic attention maps [4]or attached to fixed locations such as grids[5,3,6].