Mechanism Research on Adaptive Primitive Optimization for 3D Gaussian Splatting

With the rapid advancement of 3D vision, virtual reality, and content generation technologies, 3D scene editing has emerged as a critical task for intelligent content creation, interactive reconstruction, and spatial perception. This task imposes stringent demands on both the geometric fidelity and structural controllability of 3D models. While 3D Gaussian Splatting (3DGS) has demonstrated remarkable capabilities in novel view synthesis, it remains limited in high-precision surface reconstruction and editable scene modeling due to issues such as boundary blurring, normal discontinuities, and structural inconsistencies. To address these limitations, we propose an adaptive primitive optimization strategy tailored for geometry-aware reconstruction and structure-aware editing. Our approach refines the spatial position, anisotropic scale, and orientation of primitives through a set of targeted mechanisms, and integrates these into a unified, multi-stage training process within an end-to-end optimization framework. The resulting system establishes a closed-loop pipeline for high-quality 3D modeling, providing structural reliability and editable representation for downstream applications in scene authoring and content manipulation.