Researchers improve trampoline pose estimation by 19.6% using synthetic dataset generated from motion capture data

State-of-the-art pose estimation models struggle with extreme poses and unusual camera angles in trampoline gymnastics

Scientists created a synthetic trampoline poses (STP) dataset from motion capture recordings and realistic multiview images

Fine-tuned ViTPose model on synthetic data achieved state-of-the-art 2D results and 12.5mm 3D accuracy improvement on real trampoline footage

Pipeline converts noisy motion capture data into parametric human models to generate realistic training images from multiple viewpoints