A Method for Low-resolution Human Pose Estimation Based on Pose Distillation

The lack of detailed information in low-resolution images makes it challenging to recognize the human body structure and key points, particularly in scenarios with meager resolution. This hinders the ability of existing models to extract effective human pose features. To address this issue, this paper constructs a novel posture distiller and introduces knowledge distillation and multiscale feature fusion mechanisms. These features are then distilled to the student network, thereby compensating for the deficiency in detail information extraction observed in low-resolution scenes. Extensive comparison and ablation experiments were conducted on the COCO dataset, and the experimental results demonstrate that the proposed method enhances the AP values by 0.3% to 7.82% and 1.11% to 10.55% relative to 68.2 and 48.51 in low-resolution and very low-resolution scenarios. The model displays high accuracy across various low-resolution scenarios and exhibits enhanced pose-awareness capability, enabling it to adapt to more complex application scenarios.