Dual-Level Structural Information Learning Neural Network for Monocular 2D Pose Estimation
Zhiwen Zhang, Сонглин Ду, Dingli Luo, Takeshi Ikenaga · 2022
In this paper, a dual-level structural information learning neural network is proposed for human pose estimation referring to the human body structure. Human poses are highly structured in 2D space; thus, the structural information among human body parts is meaningful for human pose estimation. However, most existing methods do not incorporate structural information, which limits the achievable accuracy. A bidirectional memory-augmented RNN (biLSTM) based dual-level structural information learning module is proposed to compose a structural information learning neural network in which human body parts are taken as sequential information. Keypoint heatmaps and bone heat-maps are predicted through an end-to-end network, and structural information is captured at both the joint and the bone levels by exploiting the coarse keypoint heatmaps and bone heatmaps as mutual priors. Experimental results on two challenging datasets demonstrate that the proposed method achieves a PCK detection rate improvement of 0.3% on LSP dataset compared with a state-of-the-art baseline, and a PCKh detection rate improvement of 0.2% on MPII dataset over the sate-of-the-art baseline work.