Human pose estimation based on heatmap regression has achieved significant success in recent years. However, the semantic ambiguity caused by traditional hand-crafted heatmaps seriously affects the model performance. Specifically, hand-crafted heatmaps generated with a fixed Gaussian kernel are semantically misaligned. Various Gaussian covered areas for keypoints with the same type may cause model learning confusion. In this paper, we focus on learnable heatmap generation and propose a refined heatmap generator (RHG) to boost human pose estimation. First, we propose a joint training framework to connect the human pose estimator and RHG for end-to-end training. It employs a joint loss function to learn intermediate representations of the network and dataset. Second, RHG takes annotated dotpoints as input and utilizes scale-aware heatmaps as regression targets to deal with the scale variation. Scale-aware heatmaps are generated by adjusting Gaussian covered areas with geometric priors. Experimental results show that our method achieves 72.0%AP on COCO test-dev2017 and 74.0%AP on CrowdPose dataset, respectively, outperforming state-of-the-art methods.