Direct Conditional Parameterization for N-Dimensional Splatting

Key points

• Each primitive is lifted into a joint distribution over position and conditioning variables, then conditionally sliced at each rendering query.1
• We introduce direct conditional parameterization, replacing this covariance-derived readout with explicit per-effect parameters: a Cholesky precision factor for opacity and a factorized query-to-position regression operator with learnable per-dimension coupling.1
• Unrestricted conditional coordinates recover a valid joint covariance; our factorization adds spatial-scale and coupling priors.1
• Direct conditioning provides a structured, kernel-agnostic alternative for N-dimensional splatting.1

Sources (1)

• [1]Direct Conditional Parameterization for N-Dimensional Splatting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 12:39 AM

  N-dimensional splatting extends 3D Gaussian Splatting with view direction and time to model view-dependent and dynamic effects.

  Each primitive is lifted into a joint distribution over position and conditioning variables, then conditionally sliced at each rendering query.

Extractive summary: sentences quoted from the sources.