高级检索

基于结构化 4DGS 模型的单目动态场景重建

Monocular Dynamic Scene Reconstruction Based on Structured-4DGS

  • 摘要:
    针对单目动态场景重建中隐式神经辐射场方法计算复杂、渲染效率低, 以及显式3D高斯溅射重建方法在动态模糊与结构一致性方面的不足, 提出一种基于结构化4D高斯溅射的单目动态场景重建方法. 所提方法首先使用初始化点云生成锚点, 指导局部3D高斯分布建模, 根据多层级的锚点特征、视锥内的观察方向与观察距离, 实时预测局部神经高斯函数属性; 使用锚点增长和修剪细化策略, 实现高效、高质量的动态场景渲染; 最后提出一种退火平滑训练机制, 在不增加模型计算复杂度的情况下提升重建场景帧间一致性与动态细节保真度. 在D-NeRF合成数据集与NeRF-DS真实数据集上进行实验, 所提方法相较主流的Deformer-GS方法, 在峰值信噪比方面平均提升约2.7%和3.3%, 结构相似度平均提升约0.5%和4.1%, 验证了其在重建质量、结构一致性与跨视角鲁棒性等方面表现出更优性能.

     

    Abstract: Monocular dynamic scene reconstruction methods based on implicit neural radiance field models suffer from high computational cost and low rendering efficiency, while explicit 3D Gaussian splatting reconstruction approaches often introduce motion blur and fail to maintain structural consistency. To address these limitations, a structured 4D Gaussian splatting framework for monocular dynamic scene reconstruction is proposed. Multi-scale point-cloud initialization produces anchor points that guide local Gaussian modeling, and Gaussian attributes—density, color, and covariance—are predicted in real time from anchor features, viewing direction, and viewing distance; an adaptive anchor growth and pruning strategy continuously refines the Gaussian set, ensuring efficient rendering without quality loss; finally, an annealing-based smoothing mechanism enhances temporal coherence and detail fidelity without increasing model complexity. Experiments on the D-NeRF synthetic dataset and the NeRF-DS real dataset demonstrate that, compared to Deformer-GS, the proposed method achieves PSNR improvements of 2.7% and 3.3%, and SSIM gains of 0.5% and 4.1%, respectively, demonstrating superior reconstruction accuracy, structural consistency, and cross-view robustness.

     

/

返回文章
返回