Garment-aware gaussian for clothed human modeling from monocular video

<p>Reconstructing the human body from monocular video input presents significant challenges, including a limited field of view and difficulty in capturing non-rigid deformations, such as those associated with clothing and pose variations. These challenges often compromise motion editability and rendering quality. To address these issues, we propose a cloth-aware 3D Gaussian splatting approach that leverages the strengths of 2D convolutional neural networks (CNNs) and 3D Gaussian splatting for high-quality human body reconstruction from monocular video. Our method parameterizes 3D Gaussians anchored to a human template to generate posed position maps that capture pose-dependent non-rigid deformations. Additionally, we introduce Learnable Cloth Features, which are pixel-aligned with the posed position maps to address cloth-related deformations. By jointly modeling cloth and pose-dependent deformations, along with compact, optimizable linear blend skinning (LBS) weights, our approach significantly enhances the quality of monocular 3D human reconstructions. We also incorporate carefully designed regularization techniques for the Gaussians, improving the generalization capability of our model. Experimental results demonstrate that our method outperforms state-of-the-art techniques for animatable avatar reconstruction from monocular inputs, delivering superior performance in both reconstruction fidelity and rendering quality.<strong></strong></p>