Since human pose can be naturally represented by a graph, graph convolutional networks (GCNs) have recently been proposed for 3D human pose estimation and achieved promising results. But most GCN-based methods use vanilla graph convolution which aggregates features of 1-hop neighbors and long-range dependencies between joints can only be captured by stacking multiple layers of graph convolution. To alleviate this problem, we propose a multi-scale graph convolution to aggregate features of neighbors at different distances and apply it to nodes with specified neighbor types. We further propose a hierarchical-body-pooling to aggregate and share body-level and body-part-level context information. Based on these components, we finally develop a light-weighted GCN for 3D pose lifting by repeatedly stacking a residual block of multi-scale graph convolution and a hierarchical-body-pooling layer. The experimental results on Human3.6M dataset indicate that our network can achieve state-of-the-art performance with much less model complexity.

由于人体姿势可以自然地由图表示，因此最近提出了图卷积网络（GCN）用于 3D 人体姿势估计并取得了可喜的成果。但是大多数基于 GCN 的方法使用普通图卷积，它聚合了 1 跳邻居的特征，并且关节之间的远程依赖性只能通过堆叠多层图卷积来捕获。为了缓解这个问题，我们提出了一种多尺度图卷积来聚合不同距离的邻居特征并将其应用于具有指定邻居类型的节点。我们进一步提出了一个分层身体池来聚合和共享身体级别和身体部位级别的上下文信息。基于这些组件，我们最终通过重复堆叠多尺度图卷积的残差块和分层体池层，开发了用于 3D 姿态提升的轻量级 GCN。在 Human3.6M 数据集上的实验结果表明，我们的网络可以以更低的模型复杂度实现最先进的性能。