GenHMR: A Breakthrough in 3D Human Pose and Shape Estimation – Accepted at AAAI 2025
Published Feb 15, 2025 · Research
GenHMR sets a new benchmark in 3D human pose estimation by leveraging generative modeling to capture uncertainty and diversity in predictions.
Background
3D human pose and shape estimation is critical in animation, healthcare, and sports. Traditional models rely on deterministic predictions that fail in occlusion-heavy settings.
Methods
GenHMR uses a diffusion-based decoder conditioned on 2D keypoints and silhouettes. It leverages the SMPL model for realistic human shapes.
- Conditioned on OpenPose 2D keypoints.
- Trained with a large synthetic dataset.
- Incorporates a VAE encoder for latent diversity.
“This changes the game for real-time, reliable human modeling in the wild.”
Results and Impact
The model outperforms previous baselines on 3DPW and Human3.6M datasets. It also shows promising transfer performance on real-world sports video data.
Comments
Loading…