Global Attention Transformers for Extrinsic-free 3D Human Pose Classification in Squat and Deadlift

Multi-view 3D human pose estimation drives automated sports and biomechanics evaluation but relies on fragile camera calibration. Conversely, recent parameter-free alternatives degrade severely under limited views. To overcome this, we propose the Global Attention Transformer (GAT). Our method emphasizes an extrinsic-free, single-frame architecture using joint-specific queries to model spatial and cross-view relationships without geometric triangulation. We employ a synthetic-to-real training strategy using extensive viewpoint randomization for robust generalization to uncalibrated setups. We also introduce an end-to-end biomechanical assessment pipeline extracting 2D keypoints via HRNet, lifting them to 3D via GAT, and evaluating weight training movements (correct or incorrect of squats and deadlifts) using a MiniROCKET classifier. GAT achieves a competitive Mean Per Joint Position Error (MPJPE) of 68.5mm. Stress tests reveal strong generalization, maintaining 77.6mm error under monocular views. Ultimately, the end-to-end pipeline achieves 95.31 classification accuracy on our collected dataset, indicating that GAT is a promising component for camera-agnostic weight training assessment under the evaluated conditions.