A Survey on 3D Gaussian Splatting(4)
Introduction
The advent of NeRF marked a significant milestone in the landscape of computer graphics and 3D scene reconstruction, revolutionizing the way we approach novel-view synthesis.
Grounded in deep learning and computer vision , NeRF has enabled the rendering of photorealistic scenes from a sparse set of input views, establishing a new paradigm in image synthesis.
However, as with any burgeoning technology, NeRF has encountered its share of challenges and limitations , particularly in terms of computational efficiency and controllability.
It is in this context that 3DGS emerges, not merely as an incremental improvement but as a paradigm-shifting approach that redefines the boundaries of scene representation and rendering.
The journey of novel-view synthesis began long before the introduction of NeRF, with early endeavors focusing on l ight fields and basic scene reconstruction methods.
These initial techniques, however, were limited by their reliance on dense sampling and structured capture, leading to significant challenges in handling complex scenes and lighting conditions.
The emergence of structure-frommotion (SfM) and subsequent advancements in multiview stereo (MVS) algorithms provided a more robust framework for 3D scene reconstruction, setting the stage for more sophisticated view synthesis algorithms.
NeRF represents a quantum leap in this progression. By leveraging neural networks, NeRF enabled the mapping of spatial coordinates to color and density. The success of NeRF hinged on its ability to create continuous, volumetric scene function , producing results with unprecedented detail and realism.
However, this implementation came at a cost: NeRF methods were computationally intensive, often requiring extensive training times and substantial resources for rendering, especially for high-resolution outputs.
3DGS emerged as a response to these challenges. While NeRF excelled in creating photorealistic images, the need for faster, more efficient rendering methods was becoming increasingly apparent, especially for applications requiring real-time performance.
3DGS addressed this need by introducing a novel scene representation technique using millions of 3D Gaussians.
Unlike the implicit, coordinate-based models, 3DGS employs an explicit representation and highly parallelized workflows, facilitating more efficient computation and rendering.
The innovation of 3DGS lies in its unique blend of the benefits of differentiable pipelines and point-based rendering techniques.
By representing scenes with learnable 3D Gaussians, it preserves the desirable properties of continuous volumetric radiance fields, essential for high-quality image synthesis, while simultaneously avoiding the computational overhead associated with rendering in empty space, a common drawback in traditional NeRF methods.
The introduction of 3DGS is not just a technical advancement; it represents a fundamental shift in how we approach scene representation and rendering in computer graphics.
By enabling real-time rendering capabilities without compromising on visual quality, 3DGS opens up a plethora of possibilities for** applications ranging** from virtual reality and augmented reality to real-time cinematic rendering and beyond.
This technology holds the promise of not only enhancing existing applications but also enabling new ones that were previously unfeasible due to computational constraints.
Furthermore, 3DGS’s explicit scene representation offers unprecedented control over scene dynamics, a crucial factor in complex scenarios involving intricate geometries and varying lighting conditions.
This level of control and editability, combined with the efficiency of the rendering process, positions 3DGS as a transformative force in shaping future developments in the related field.