MPA-GNet: multi-scale parallel adaptive graph network for 3D human pose estimation

Graph convolutional networks (GCNs) have achieved remarkable performance in the 2D-to-3D human pose estimation (HPE) task. The adjacency matrix in GCNs is crucial for feature aggregation in 3D HPE. However, existing GCN-based methods excessively rely on the fixed adjacency matrix to aggregate joint features from one-hop neighbor at a single scale, which limits the feature representation of skeleton data. To better improve the performance of 3D HPE, we have designed a multi-scale parallel adaptive graph network (MPA-GNet) for 3D HPE. The proposed network consists of three parallel multi-scale subgraph networks (PMS-Net) to efficiently capture human joint features at different scales. Specially, a multi-scale feature fusion module is devised to process multi-scale graph structural features and exchange information to generate rich hierarchical representations for skeleton data. To flexible construct graph topology in different scales, a special designed adaptive attention adjacency graph convolution network and a cluster graph pooling module are designed to construct the MPA-GNet in a parallel manner and capture the local subgraphs information in each PMS-Net. Finally, we conduct experiments on two 3D human pose challenging benchmark datasets Human3.6M and HumanEva-I for evaluating the effectiveness of the proposed model. The experimental results demonstrate that our model achieves competitive performance compared with some state-of-the-art 3D HPE methods.