Human Pose Estimation Method Based on Spatial-Aware Embedding

In multi-person pose estimation tasks, keypoint grouping methods based on embedded labels are widely used. However, in complex scenarios such as occlusion and crowding, the results of keypoint grouping are often suboptimal. To address these challenges, this paper proposes a spatial-aware labeling method for multi-person pose estimation. During label initialization, different initial values are assigned to the labels based on the body part to which the keypoint belongs. Additionally, a coordinate encoding method is designed to transform keypoint coordinates into numerical vectors embedded within the labels, forming multi-dimensional labels. The labels generated by this method carry spatial information from the image, enhancing the model's spatial perception. A corresponding loss function is designed to preserve these features during model training. The effectiveness of this method is validated by testing the model's performance across various scenarios on the COCO dataset.