SkelSplat: Robust Multi-view 3D Human Pose Estimation with Differentiable Gaussian Rendering

The introduction of SkelSplat marks a significant advancement in multi-view 3D human pose estimation, addressing the limitations of existing methods that struggle with generalization across varying test scenarios. Traditional approaches rely heavily on large annotated datasets, which can lead to poor performance when conditions change. SkelSplat leverages differentiable Gaussian rendering to model human poses as a skeleton of 3D Gaussians, allowing for seamless integration of multiple camera views without the need for 3D ground-truth supervision. This innovative framework has shown to outperform existing methods in benchmarks like Human3.6M and CMU, achieving a remarkable reduction in cross-dataset error by up to 47.8%. Furthermore, experiments conducted on datasets such as Human3.6M-Occ and Occlusion-Person demonstrate its robustness to occlusions, making it a versatile solution for real-world applications in augmented reality and human-robot interaction.