Efficient Low-Resolution Human Pose Estimation with Knowledge Bridge

In the field of human pose estimation, the problem of knowledge transfer among models with different learning objectives and input resolutions remains unresolved. In this paper, we propose a novel framework specifically designed for low-resolution human pose estimation, aiming to obtain an efficient and powerful model for this task. Specifically, we introduce an intermediate model as a knowledge bridge between the teacher model and the student model. We perform knowledge transfer from the teacher model with different learning objectives and resolutions to the student model through cross-learning-objective pose distillation between the teacher model and the bridge model, as well as cross-resolution pose distillation between the bridge model and the student model. Experimental results on the COCO dataset demonstrate the effectiveness of our model in low-resolution human pose estimation tasks.