计算机科学
人工智能
稳健性(进化)
计算机视觉
渲染(计算机图形)
压扁
高斯分布
高斯过程
水准点(测量)
曲面重建
最优化问题
视图合成
三维重建
高斯网络模型
迭代重建
计算机图形学(图像)
三维模型
出处
期刊:
[SPIE]
日期:2025-09-19
卷期号:: 166-166
被引量:2
摘要
To address the lack of geometric constraints in existing 3D Gaussian Splatting (3DGS) methods, we propose a novel geometry-aware optimization framework, ANormGS, which integrates anchor guidance with surface normal priors. To mitigate local distortions, Gaussian redundancy, and rendering artifacts caused by randomly oriented ellipsoids, ANormGS incorporates normal supervision into its anchor-based splatting pipeline. First, a dual optimization strategy is introduced, which combines normal-aware rendering with pretrained surface normals to align Gaussian orientations with underlying scene geometry. Second, we propose a scale-constrained flattening approach to enhance geometric compactness and suppress redundant representations. Extensive experiments on benchmark datasets, including Tanks and Temples, Deep Blending, and Mip-NeRF 360, demonstrate that ANormGS outperforms baseline methods, achieving an average PSNR gain of 0.14 dB, an SSIM improvement of 0.007, and an LPIPS reduction of 11.6%. These results highlight the effectiveness and robustness of our method across diverse and challenging 3D reconstruction scenarios.
科研通智能强力驱动
Strongly Powered by AbleSci AI