CODN-GS: Coupled Optimization of Depth and Normal in 3D Gaussian Splatting for Scene Reconstruction

Scene reconstruction has attracted widespread attention due to its extensive applications in intelligent devices. Recently, 3D Gaussian Splatting (3DGS) has gained recognition as a prominent technique owing to its impressive pixel-level rendering quality and high reconstruction speed. However, 3DGS scenes constructed with only photometric supervision are highly prone to RGB overfitting. This problem neglects the depth of objects and the surface normal in the reconstructed scene, which leads to the lack of geometric consistency. To address these issues, this study introduces a coupled optimization of depth and normal within the 3D Gaussian Splatting (CODN-GS) framework. Firstly, a normal–depth–normal transformation is applied to ensure accurate capture of geometric features in the reconstructed scenes. Secondly, a robust monocular depth supervision model generates depth maps that are refined via global and local adjustments, which serve as guidance for the model to accurately learn scene geometry. Thirdly, a normal supervision model is incorporated as a complement to depth supervision, jointly optimizing the overall scene geometry. Finally, comprehensive experiments on the Replica, MipNerf360, and ScanNet datasets demonstrate that CODN-GS reduces RMSE-D and RMSE-N by at least 9% and 13%. These results confirm that the proposed method outperforms state-of-the-art methods in both depth and normal accuracy. Note to Practitioners—While 3D Gaussian Splatting (3DGS) has shown promise for scene reconstruction, achieving accurate geometric consistency with purely image-based methods remains challenging. Existing approaches primarily optimize 3DGS using RGB information, with some incorporating depth and normal constraints. However, these methods struggle in complex regions, such as textureless and reflective surfaces, which often lead to geometric distortion. To address these issues, we introduce CODN-GS (Coupled Optimization of Depth and Normal within 3D Gaussian Splatting), a novel reconstruction framework designed to improve geometric accuracy without the need for additional sensors. CODN-GS first applies a surface refinement transformation to improve the model’s ability to learn depth and normal information. It then derives depth and normal maps for each input RGB image using transfer learning and fine-tuning, providing robust geometric supervision for reconstruction. We evaluated CODN-GS on three public datasets and found that it significantly improves geometric consistency while maintaining high photometric reconstruction accuracy.