Unbiased feature position alignment for human pose estimation

Multi-scale feature fusion is a commonly-used module in existing deep-learning models, and feature misalignment occurs in the process of feature fusion. The spatial misalignment hinders the learning of semantic representation with multi-scale levels, but which has not received much attention. This misalignment problem is caused by the feature position shift after using the convolution and interpolation operation in feature fusion. To solve the misalignment problem, this paper formulates the shift error mathematically and proposes a plug-and-play unbiased feature position alignment strategy to align convolution with interpolation. As a model-agnostic approach, unbiased feature position alignment can boost the performance of different models without introducing extra parameters. Furthermore, the unbiased feature position alignment is applied to build an unbiased human pose estimation method. Experimental results have demonstrated the effectiveness of the proposed unbiased pose model in comparison to the state-of-the-arts, especially in the low-resolution field. The codes are shared at https://github.com/WangChen100/Unbiased-Feature-Position-Alignment-for-Human-Pose-Estimation.