Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction
Optimizes 3D Gaussians in canonical space with a deformation field for monocular dynamics and uses annealed smoothing to reduce temporal jitter from pose noise.
Authors / Team
Ziyi Yang · Researcher
Year
2024
Deep Dive
The approach extends static 3DGS to monocular dynamic scenes by coupling learnable Gaussians in canonical space with an implicit deformation field over time. A custom differentiable Gaussian rasterizer supplies gradients for both components. An annealing smoothing training schedule mitigates pose inaccuracies without extra runtime cost. Experiments report improved rendering quality and real-time speed versus implicit dynamic NeRF-style baselines on standard benchmarks.
What we learn
- 01
Canonical plus deformation remains the backbone for extending explicit points to dynamic video.
- 02
Pose-sensitive losses benefit temporal regularizers or annealing when cameras are imperfect.
Verbatim quote
"Implicit neural representation has paved the way for new approaches to dynamic scene reconstruction and rendering."— source ↗
Tags
Links
Sources