Wang Xinshun ; Cui Qiongjie ; Chen Chen ; Liu Mengyuan

Abstract

The past few years has witnessed the dominance of Graph ConvolutionalNetworks (GCNs) over human motion prediction.Various styles of graphconvolutions have been proposed, with each one meticulously designed andincorporated into a carefully-crafted network architecture. This paper breaksthe limits of existing knowledge by proposing Universal Graph Convolution(UniGC), a novel graph convolution concept that re-conceptualizes differentgraph convolutions as its special cases. Leveraging UniGC on network-level, wepropose GCNext, a novel GCN-building paradigm that dynamically determines thebest-fitting graph convolutions both sample-wise and layer-wise. GCNext offersmultiple use cases, including training a new GCN from scratch or refining apreexisting GCN. Experiments on Human3.6M, AMASS, and 3DPW datasets show that,by incorporating unique module-to-network designs, GCNext yields up to 9x lowercomputational cost than existing GCN methods, on top of achievingstate-of-the-art performance.