Real-time reconstruction frame rate of landscape display scene optimized by 3D Gaussian sputtering
Abstract
The real-time reconstruction of landscape exhibition scenes puts forward high requirements for image continuity, viewpoint switching and interactive response. Aiming at multi-view data redundancy, complex spatial structure and rendering load fluctuation, this paper proposes a real-time reconstruction frame rate optimization method based on 3D Gaussian Splatting. The method takes multi-view images, camera pose, depth information and interactive trajectory as inputs, and controls Gaussian density and level of detail through keyframe screening, feature encoding, Gaussian point cloud initialization and lightweight representation. On this basis, parameter update, visibility culling, tiled rendering and cache reuse are combined to adjust computational overhead and maintain stable rendering during continuous interaction. Four representative scenes, including outdoor garden exhibition, indoor landscape sand table, night lighting exhibition and mixed installation exhibition, are used to verify the reconstruction performance. The experimental results show that the PSNR of the optimized method reaches 32.7, 33.4, 32.0 and 32.6 dB respectively, and the SSIM remains above 0.931. The LPIPS is controlled within 0.082, and the reconstruction error remains within 0.92 - 1.03 cm, indicating that the proposed method does not weaken key visual details while optimizing frame rate. The FPS, response delay and frame rate stability scores reach 91.8, 90.7 and 89.5, respectively. This study can provide a method reference for real-time reconstruction optimization in digital landscape exhibition and virtual interactive scenes.