3D Gaussian Splatting (3DGS) represents a significant advancement in neural rendering, offering high-quality, real-time novel view synthesis by modeling 3D scenes with explicit 3D Gaussians rendered via rasterization. This technique provides substantial improvements in training speed and rendering performance over previous methods like NeRF, enabling a wide array of applications from dynamic scene reconstruction and human avatars to generative AI tasks and SLAM.

How It Works

The core of 3DGS involves representing a scene as a collection of 3D Gaussians, each defined by its position, covariance (shape and orientation), color, opacity, and spherical harmonics for view-dependent appearance. These Gaussians are efficiently rendered using a rasterization pipeline, which projects them onto the 2D image plane. Key research directions focus on optimizing Gaussian attributes, managing their density through adaptive splitting and pruning, and developing efficient rendering algorithms that support anisotropic splatting and visibility awareness.

Quick Start & Requirements

This document is a survey of research papers on 3D Gaussian Splatting, not a single project repository. Specific implementation requirements vary, but generally include a Python environment with deep learning libraries such as PyTorch or TensorFlow, and a CUDA-enabled GPU for efficient training and rendering. Many papers provide links to their code repositories for detailed setup instructions.

Highlighted Details

• Efficiency & Compression: Significant research focuses on reducing memory footprint and storage through techniques like vector quantization, pruning, and hierarchical representations (e.g., LightGaussian, ContextGS, F-3DGS), enabling efficient streaming and deployment.
• Real-time Performance: Achieves high-quality rendering at resolutions like 1080p and high frame rates (e.g., 200+ FPS), often surpassing previous neural rendering methods.
• Generative AI Integration: Seamlessly combined with diffusion models for text-to-3D content creation, scene editing, and stylization, often leveraging techniques like Score Distillation Sampling (SDS).
• Diverse Applications: Extended to dynamic scenes, animatable human avatars, SLAM systems, autonomous driving, and detailed object reconstruction from sparse views.
• Improved Reconstruction & Regularization: Advancements in initialization strategies, geometry-aware regularization (e.g., depth, SDF guidance), and surface extraction methods enhance reconstruction accuracy and detail.

Maintenance & Community

As this document is a survey of research papers rather than a single software project, there is no central maintenance or community in the traditional sense. However, the rapid pace of publication and the active research community indicate strong ongoing development and interest in the field.

Licensing & Compatibility

Licensing information is specific to each individual research paper and its associated code repository. Users should consult the respective licenses for each implementation they intend to use. Compatibility for commercial use or closed-source linking will vary.

Limitations & Caveats

While 3DGS has made significant strides, limitations persist. Some methods may still struggle with highly reflective or transparent surfaces, texture-less regions, and precise surface reconstruction without specialized regularization. Handling extreme dynamic scenes or achieving perfect multi-view consistency in generative tasks remains an active area of research. The field is rapidly evolving, with new techniques and improvements emerging frequently.