DiffPose: Reliable 2D Pose Estimation Through Denoising Diffusion

We present DiffPose, a novel framework for 2D human pose estimation. DiffPose is capable of generating reliable lower-uncertainty heatmap from noise using a given image. DiffPose differs from previous methods in that it corrects the deviation in its own predictions without designing additional pose refinement modules. To accomplish these, we introduce a forward diffusion process that converts the heatmap into random noise. We then employ iterative denoising to generate reliable pose estimation during the reverse diffusion process. Extensive experiments on the well-recognized MS-COCO benchmark demonstrate the superior reliability of DiffPose compared to previously well-established detectors.