This paper tackles the problem of human motion prediction, consisting in forecasting future body poses from historically observed sequences. Despite of their performance, current state-of-the-art approaches rely on deep learning architectures of arbitrary complexity, such as Recurrent Neural Networks~(RNN), Transformers or Graph Convolutional Networks~(GCN), typically requiring multiple training stages and more than 3 million of parameters. In this paper we show that the performance of these approaches can be surpassed by a light-weight and purely MLP architecture with only 0.14M parameters when appropriately combined with several standard practices such as representing the body pose with Discrete Cosine Transform (DCT), predicting residual displacement of joints and optimizing velocity as an auxiliary loss. An exhaustive evaluation on Human3.6M, AMASS and 3DPW datasets shows that our method, which we dub siMLPe, consistently outperforms all other approaches. We hope that our simple method could serve a strong baseline to the community and allow re-thinking the problem of human motion prediction and whether current benchmarks do really need intricate architectural designs. Our code is available at \url{https://github.com/dulucas/siMLPe}.

本研究提出了一种轻量级的人体运动预测网络“siMLPe”，结合了多层感知器、离散余弦变换、关节残差位移预测和速度辅助损失优化等简单方法，且具有仅有0.14万个参数和超过三个基准数据集的卓越表现，成为该领域的强基准。

回到 MLP：人体动作预测的简单基线