DDP-Gaussian: dense depth priors guide Gaussians optimization

3D Gaussian Splatting (3D-GS) has achieved a breakthrough in the field of 3D reconstruction through the use of differentiable rasterization of 3D Gaussian primitives. Compared to Neural Radiance Fields (NeRF), 3D-GS exhibits superior rendering efficiency and quality, significantly accelerating training speed and enabling real-time high-quality rendering. The initialization of Gaussians is critical to reconstruction results. The original 3D-GS utilizes the output of the sparse reconstruction as input and dynamically generates Gaussians via an adaptive density control strategy, compensating for the limitations of sparse points. However, during optimization, 3D-GS is prone to local optima, leading to artifacts and distortions.Furthermore, the adaptive density control strategy of the original 3D-GS does not adequately focus on fine texture structures, thus limiting further reconstruction enhancement. To address these issues, we propose a Gaussian reconstruction method based on dense depth priors. We first leverage depth prior information to generate a sufficient number of initial Gaussians and then apply an adaptive density optimization algorithm to refine texture-rich regions, achieving superior reconstruction results. Experiments conducted across various scenes demonstrate that our method outperforms existing state-of-the-art algorithms across multiple evaluation metrics.