Abstract
Neural rendering methods represent a crucial and challenging task in
computer vision. Recent advancements in 3D Gaussian Splatting have set
new benchmarks in rendering quality and speed by combining the strengths
of geometric primitives and volumetric representations. In contrast to
the point-based approach of 3D Gaussian Splatting, Scaffold-GS obtains
anchors to achieve significant memory reduction. In this paper, we use
Scaffold-GS as the baseline to achieve high quality rendering with
minimal anchors. We propose a depth guide structured 3D Gaussian method
for real-time rendering, called DS-GS. Our method employs depth priors
to guide two types of anchor growth, ensuring that anchors develop in
the correct positions during training. And we implement a differentiable
depth rasterizer to enhance the similarity between rendered depth and
estimated depth. Additionally, we use multi-scale training to improve
rendering quality. We demonstrate the advantages and potential of our
approach across a variety of scenarios, effectively reducing anchor
redundancy and improving rendering quality. For intance, in the
Mip-NeRF360 dataset, we increased the PSNR by 0.6 and reduced the
storage size by 30%. Furthermore, our approach shows robustness in
experiments with real-world scenes.