A Head Pose Estimation Method Based on Improved RepVgg Network

Head pose estimation refers to the task of inferring the spatial rotation of the human head relative to the camera coordinate system. Accurate head pose estimation is of great importance in various fields such as behavior understanding, emotion recognition, and human-computer interaction. In this paper, we propose a head pose estimation algorithm based on RepVGG network. Unlike methods that represent head pose using Euler angles or quaternions, our approach adopts rotation matrices for pose representation and employs a geodesic loss function to train the model. To enhance the feature extraction capability of the network, we incorporate SE attention mechanisms at different stages of the RepVGG architecture. We evaluate the effectiveness of our approach on the public AFLW2000 dataset. Experimental results show that these improvements reduce the mean absolute error and positively contribute to the model's accuracy.