DCARE: Dynamic Clothed Avatar Reconstruction With Normal-Aware Gaussian Splatting From Monocular Videos

Reconstructing dynamic clothed avatars from monocular video using 3D Gaussian Splatting (3DGS) remains challenging due to complex clothing deformations and inconsistent appearances. We present DCARE, an integrated framework that addresses these issues by directly leveraging surface curvature, extracted from normal maps, to guide Gaussian density control. At the core of our approach is a curvature-guided Adaptive Density Control (ADC) mechanism that allocates Gaussian density dynamically according to local geometric complexity. This design differs fundamentally from prior normal-based 3DGS methods, which rely on normals primarily for regularization losses rather than direct density management. To capture motion, DCARE introduces a synergistic deformation model that combines adaptively determined virtual bones for large-scale movements with normal-guided nonrigid refinement for fine-grained details. For appearance, consistency is maintained via a global VQ-VAE codebook, which provides a robust prior decoupled from geometry. As a result, DCARE achieves state-of-the-art performance in dynamic clothing reconstruction, particularly excelling in scenarios with loose garments and semi-transparent textures, significantly improving both geometric accuracy and visual fidelity.