Diffusion-Refinement Pose Estimation With Hybrid Representation

2-D human pose estimation is a measurement means based on machine vision, which aims to locate the 2-D coordinates of body joints in images. According to the supervisory signals adopted, existing pose estimation studies can be broadly divided into heatmap-based methods (accurate but not end-to-end) and regression-based methods (end-to-end but not accurate). There is obviously a complementarity between these two pose estimation paradigms, but it has not received enough attention. In this article, a novel hybrid-representation model hybrid representation pose estimation (HRPE) is proposed to deeply correlate heatmap representation and coordinate regression via a truncated diffusion model. Specifically, HRPE mainly consists of two parts: 1) preliminary prediction adopts a pose estimation model based on heatmap representation to get a robust prior for initialization and 2) iterative refinement performs a reverse denoising process based on coordinate regression for stepwise correction. Although truncated diffusion models are common in generation tasks, their application to perception tasks is not straightforward due to the differences in the objectives of the two types of tasks. In this regard, a qualitative analysis is provided on the rationality of truncation in this article, which is not done in other perception works using truncated diffusion models. In summary, a superior strategy is proposed to combine heatmap representation and coordinate regression to exploit their complementarity for better performance. Experimental results show that our hybrid-representation model HRPE achieves 77.9 and 76.7 AP on the COCO validation and test-dev sets.