Computer Graphics

Neural Representations

Learned scene representations, and their reconstruction lineage.

• supersedescorrects · extends
• classifiesspecializes · part-of
• substitutes forapproximates · alternative-to
• depends onrequires · validates

Colour is the family; a dashed line is the second member of it.

The 12 reasons on these edges, as text

• alternative-to3D Gaussian Splatting is a competing approach to Neural Radiance Fieldsrasterizes explicit primitives instead of querying a field
• requires3D Gaussian Splatting does not work without Rasterizationthe speed comes from sorted alpha-blended splatting
• requires3D Gaussian Splatting does not work without Stochastic Gradient Descentthe primitives are fitted to photographs by gradient descent
• requiresInstant NGP does not work without Multiresolution Hash Encodingthe speedup is the encoding, not the network
• correctsMip-NeRF fixes a defect in Neural Radiance Fieldspoint samples alias when rendered at another scale
• correctsMultiresolution Hash Encoding fixes a defect in Neural Radiance Fieldsthe deep MLP, not the idea, is what made it slow
• extendsNeural Radiance Cache adds capability to Path Tracingterminates paths into a cache learned while rendering
• alternative-toNeural Radiance Cache is a competing approach to Irradiance Cachinga learned cache instead of an interpolated one
• requiresNeural Radiance Cache does not work without Stochastic Gradient Descentthe cache is trained online, while rendering
• specializesNeural Radiance Fields is a specific case of Radiative Transfer Equationan MLP supplies density and colour along the ray
• requiresNeural Radiance Fields does not work without Stochastic Gradient Descentthe field is fitted to photographs by gradient descent
• correctsPlenoxels fixes a defect in Neural Radiance Fieldsa sparse grid works too; the network was not the point

7 nodes