search

ingra14m / Deformable-3D-Gaussians

[CVPR 2024] Official implementation of "Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction"
https://ingra14m.github.io/Deformable-Gaussians/
MIT License
779 stars 40 forks source link

How to understand this sentence in the paper? #39

Open isxulu opened 3 months ago

isxulu commented 3 months ago

Thank you for your nice work!

image image

I am sorry but How to understand this sentence in the paper? "Explicit point-based rendering further elevates the rank of the scene."

ingra14m commented 3 months ago

Hi, thanks for the interest.

The term "low-rank tensor assumption" originates from TensoRF. Structures such as K-plane, Hexplane, Tensor4D, etc., rely on this assumption. However, the rank of a scene is related to its topological structure and the degree of variation (dynamic scenes). Thus, strictly speaking, dynamic scenes do not satisfy the low-rank tensor assumption. Our early experiments have also proven this point, hence we utilized a pure MLP to model deformation.

Secondly, the essence of introducing deformation is to incorporate the constraint of motion smoothness. The grid-based structure results in non-smooth deltas, which limits its potential.