SRP-Avatar: Monocular 3D Head Avatar Reconstruction with Semi-Rigid Geometric Constraints and Pose Refinement

Abstract Recent advancements in 3D Gaussian Splatting (3DGS) have established a strong benchmark for dynamic head avatar reconstruction. However, achieving improved physical consistency and local photorealism remains challenging, as inherent tracking jitter often blurs high-frequency details, and the complex oral topology frequently leads to non-physical geometric distortions. Furthermore, standard shading networks struggle to accurately disentangle material properties for regions with distinct optical characteristics, such as eyes and teeth. To address these limitations, we propose a 3D Gaussian head avatar framework that explicitly targets pose jitter, intra-oral geometric instability, and specular degradation. We counteract tracking instability by integrating a Perceptually-Guided Adaptive Pose Refinement module. This module incorporates LPIPS-driven gradients into the pose optimization loop, enhancing texture sharpness and reducing noise-induced blur. It does so by refining the pose parameters based on perceptual quality, to achieve more accurate and stable head geometry. Addressing anatomical plausibility, we redesign the deformation field with an Anatomically-Adaptive Semi-Rigid Constraint, coupling rigid mandibular transformations with soft skinning to mitigate geometric collapse in the oral cavity. Complementing these geometric improvements, a Semantically-Aware Specular Injection mechanism is introduced during the rendering stage, leveraging semantic priors to recalibrate material attributes for enhanced specularity in eyes and teeth. Experiments on INSTA and HDTF show improved perceptual quality and more stable intra-oral geometry compared with prior 3DGS head avatar baselines.