3DSENet: 3D Spatial Attention Region Ensemble Network for Real-time 3D Hand Pose Estimation

摘要

The data representations of existing depth-based 3D hand pose estimation methods include depth images, volumetric representation and point cloud. Compared with the other two modalities, point cloud can directly represent the spatial information and avoid the quantization error caused by the voxelization. Furthermore, considering both the accuracy and the real time performance, we design a novel 3D deep network for point cloud, named 3D spatial attention region ensemble network (3DSENet), with smaller model, less training time and higher frame rate. Taking point cloud as input, our proposed 3DSENet is able to capture local structure information and makes fully use of the physical constraints in fingers with the guidance of an initial hand pose. The experimental results on three public datasets demonstrate that our approach helps improve the final accuracy and achieves comparable performance with state-of the-art methods.