Star Operation in Self-Attention for 3D Human Pose Estimation

Abstract

Recent transformer-based methods have achieved notable success in 3D human pose estimation. However, the most utilized self-attention mechanisms compute the attention matrix by performing a dot product on inter-vector features, which may overlook finer element-wise interactions. In this paper, we introduce star-attention, which integrates the star operation into the self-attention module. This approach retains the non-linearity and high dimensionality characteristics of matrix multiplication in self-attention while enhancing the feature representation by capturing interactions at a finer granularity. Specifically, we validated the promotion of the proposed star-attention module on MixSTE as an example. Experimental results demonstrate that our approach achieves competitive performance on benchmarks compared to state-of-the-art methods, producing smoother 3D poses across successive frames.