A Dual-Path Attention Fourier Convolutional Network for Human Motion Prediction

Human Motion Prediction is developed to project human motion in the future frames.With lots of papers tend to predict the future motion via Recurrent Neural Network, Multi-Layer Perceptron, or Graph Convolution Network, many complicated motions have been predicted more accurately.However, most existing methods have met drawbacks in long-term predictions.To make the long-term prediction better, we propose a network called DAFCN, using Fast Fourier Convolution to optimize the result on long-term human motion prediction.Our approach also makes advantages of the Motion Attention Model to abstract the short-term useful information.Furthermore, our approach attempts to make prediction in short-term and long-term paths, which will be mixed in the output layers.Currently, our approach is experimented on Human3.6M,which demonstrate a better result on prediction.The code is available at: https://github.com/Kiramei/DAFCN".