Human Pose Estimation Based on Data Augmentation and Contextual Relation Attention

Human pose estimation is a critical research area in computer vision, aiming to predict key points' spatial coordinates in images. This paper introduces a novel network for human pose estimation that addresses performance issues caused by occlusion. To combat overfitting, we use data augmentation techniques like joint occlusion and instance pasting. We also propose a Contextual Relation Attention Module (CRAM) to enhance focus on important features, reduce background noise interference, and improve accuracy by handling limb occlusion challenges. Additionally, we incorporate a Transformer for inference to predict occluded joint positions more accurately by learning long-range dependencies between joints. Experimental results on benchmark datasets demonstrate the superiority of our method over existing approaches.