Visual depth models coupled with 3D pose estimation for sports body training evaluation

Traditional methods for assessing human body aesthetics largely overlook the influence of dynamic three-dimensional postures, limiting their real-world applicability.To address this gap, we propose vision transformer-three-dimensional pose, a novel framework that integrates a vision transformer with graph convolutional networks to achieve accurate pose-aware aesthetic quantification.Our approach begins by extracting global visual features via a vision transformer and concurrently models skeletal kinematics using a graph convolutional network.These cross-modal features are then fused through a gated attention mechanism, guided by biomechanical constraints.Finally, a joint loss function co-optimises pose accuracy and aesthetic perception.Experimental results demonstrate that our method achieves a pose error of 43.7 mm on the Human3.6Mdataset and a spearman rank correlation of 0.917 on the South China University of Technology -Facial Beauty Prediction 5,500 dataset, outperforming existing state-of-the-art methods and significantly enhancing robustness under diverse viewing angles.