3D Gaussian Splatting
A scene reconstruction and real-time rendering method based on three-dimensional Gaussian representations. It achieves high-quality novel view synthesis through explicit Gaussian point clouds.
Timeline (last 90 days)
NeRF and 3D Gaussian splatting are the two current mainstream novel view synthesis methods: the former represents the radiance field implicitly with a multilayer perceptron, and the latter represents the scene explicitly with Gaussian ellipsoids.
A Gaussian representation contains multiple attributes such as position, covariance, spherical harmonic coefficients and opacity; if geometry and appearance are encoded separately, their strong physical correlations cannot be captured by a joint entropy model, resulting in wasted bitrate.
3D Gaussian Splatting (3DGS) is a 3D scene representation and rendering technology proposed in 2023, which explicitly represents scenes with millions to billions of 3D ellipsoids carrying direction, color, transparency and spatial scale, and achieves real-time rendering through rasterization.
Compared with NeRF, 3DGS trains faster and can render at real-time frame rates, but storage and VRAM usage grow linearly with the number of Gaussians, and billion-level scale places very high demands on compression, tiling and engine integration.
3D digital assets generated by 3D Gaussian scanning have the characteristics of high realism, fast production efficiency and low deployment cost, and have application space in directions such as digital protection of cultural relics and ancient buildings, immersive cloud tours and digital cultural creativity.
Compared to NeRF, 3DGS renders tens of times faster and can directly output geometric information usable by physics engines.
Splat2Act uses 3D Gaussian splatting to construct photorealistic digital twins while also generating collision geometry for physical interaction in simulation.
The difference between Splat2Act's simulation success rate and real-robot success rate is less than 3 percentage points, primarily because the 3DGS reconstruction produces visual inputs with nearly the same distribution as the real scene, fundamentally compressing the visual domain gap.
Meshes are an explicitly discretized representation that is natively compatible with mainstream game engines (such as Unity and Unreal Engine) and DCC tools (such as Blender and Maya), whereas implicit representations like NeRF or 3D Gaussian Splatting are difficult to edit locally with fine granularity.
All Facts (9)
NeRF and 3D Gaussian splatting are the two current mainstream novel view synthesis methods: the former represents the radiance field implicitly with a multilayer perceptron, and the latter represents the scene explicitly with Gaussian ellipsoids.
50%UnverifiedA Gaussian representation contains multiple attributes such as position, covariance, spherical harmonic coefficients and opacity; if geometry and appearance are encoded separately, their strong physical correlations cannot be captured by a joint entropy model, resulting in wasted bitrate.
50%Unverified3D Gaussian Splatting (3DGS) is a 3D scene representation and rendering technology proposed in 2023, which explicitly represents scenes with millions to billions of 3D ellipsoids carrying direction, color, transparency and spatial scale, and achieves real-time rendering through rasterization.
50%UnverifiedCompared with NeRF, 3DGS trains faster and can render at real-time frame rates, but storage and VRAM usage grow linearly with the number of Gaussians, and billion-level scale places very high demands on compression, tiling and engine integration.
50%Unverified3D digital assets generated by 3D Gaussian scanning have the characteristics of high realism, fast production efficiency and low deployment cost, and have application space in directions such as digital protection of cultural relics and ancient buildings, immersive cloud tours and digital cultural creativity.
50%UnverifiedCompared to NeRF, 3DGS renders tens of times faster and can directly output geometric information usable by physics engines.
50%UnverifiedSplat2Act uses 3D Gaussian splatting to construct photorealistic digital twins while also generating collision geometry for physical interaction in simulation.
50%UnverifiedThe difference between Splat2Act's simulation success rate and real-robot success rate is less than 3 percentage points, primarily because the 3DGS reconstruction produces visual inputs with nearly the same distribution as the real scene, fundamentally compressing the visual domain gap.
50%UnverifiedMeshes are an explicitly discretized representation that is natively compatible with mainstream game engines (such as Unity and Unreal Engine) and DCC tools (such as Blender and Maya), whereas implicit representations like NeRF or 3D Gaussian Splatting are difficult to edit locally with fine granularity.
50%