Multi person pose estimation based on CspNeXt and Vipnas

The paper introduces a multi-person pose estimation method utilizing the CspNeXt-t network, which substantially enhances the representation capabilities of key point features by incorporating a self-attention mechanism. We employed the TopDownHeatmap method for pose prediction and integrated the Vipnas technique for dynamic optimization of the network architecture. During the training process, we leveraged pre-trained models and implemented the cosine annealing strategy, along with training pipelines and hyperparameter iteration techniques, to enhance training efficiency and model performance. The experimental results show that this method achieves excellent performance on the COCO dataset.