More accurate heatmap generation method for human pose estimation

Human pose estimation plays a crucial role in computer vision, such as understanding body language and tracking behavior. We observed that neural networks trained to generate heatmaps of human joints often produce blurred outputs that lacks a well-defined Gaussian structure. Despite this, existing methods largely prioritize network architecture innovations, neglecting heatmap generation itself. In light of the discovered importance, we propose a novel approach that incorporates a visual center module and a heatmap enhancer to improve existing human pose estimation methods. First, we extract features from the backbone network (any model based on convolutional neural networks) at different depths. The visual center module is then used to capture the global long-range dependencies and the cross-channel message of these features, which facilitates the optimal generation of the heatmap. Finally, the heatmap value distribution is adjusted using the heatmap enhancer. The heatmap enhancer can handle multiple peaks around the maximum activation through a Gaussian filter, allowing the heatmap to achieve accurate localization of the body’s joint points. We combine our module with current mainstream human pose estimation methods for experiments. The experimental results show that the proposed method has achieved good results on the two benchmark datasets of MSCOCO and MPII.