Lightweight Stacked Hourglass Network for Efficient Robotic Arm Pose Estimation

Existing robotic arm pose estimation methods usually only focus on improving the accuracy of the model, but ignoring model efficiency, therefore the development of models with a large number of parameters and computation in practical use. To solve the problem, this paper proposes a new HetConv-Based Stacked Hourglass Network (HetConv-SHN). HetConv-SHN develops from the Stacked Hourglass Network (SHN), which constructs an SHN with two stacks, and compresses the model by replacing the standard convolution in the SHN with a lightweight HetConv (Heterogeneous Kernel-Based Convolution) module. Experiments conducted on four typical datasets demonstrate that HetConv-SHN can achieve better accuracy than its counterparts with fewer parameters and computation. It can be better deployed on robotic arms, enabling lightweight robotic arm pose estimation.