alphaXiv(@askalphaxiv):"Depth Anything V4" Depth Anything V4 reconstructs dynamic 4D scenes from monocular video using Riemannian Flow Matching over 4D Gaussian parameters. Instead of forcing rotations, scales, and opacity through Euclidean space, it follows their natural manifolds, cutting invalid Gaussians from 12.3% to 0.01% and improving F1 from 0.762 to 0.806 over an otherwise identical MLP baseline. Moving from frame-wise depth prediction to probabilistic 4D scene understanding.

2026.08.20 20:20

"Depth Anything V4" Depth Anything V4 reconstructs dynamic 4D scenes from monocular video using Riemannian Flow Matching over 4D Gaussian parameters. Instead of forcing rotations, scales, and opacity through Euclidean space, it follows their natural manifolds, cutting invalid Gaussians from 12.3% to 0.01% and improving F1 from 0.762 to 0.806 over an otherwise identical MLP baseline. Moving from frame-wise depth prediction to probabilistic 4D scene understanding.

显示更多