Monocular Dynamic Gaussian Splatting: Fast, Brittle, and Scene Complexity Rules

Authors: Yiqing Liang, Mikhail Okunev, Mikaela Angelina Uy, Runfeng Li, Leonidas Guibas, James Tompkin, Adam W Harley

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we organize, benchmark, and analyze many Gaussian-splatting-based methods, providing apples-to-apples comparisons that prior works have lacked. We use multiple existing datasets and a new instructive synthetic dataset designed to isolate factors that affect reconstruction quality. We systematically categorize Gaussian splatting methods into specific motion representation types and quantify how their differences impact performance.
Researcher Affiliation Collaboration 1Brown University 2Stanford University 3NVIDIA
Pseudocode No The paper describes methods and mathematical definitions but does not contain explicit pseudocode or algorithm blocks.
Open Source Code Yes Public release. Code and qualitative results including our instructive synthetic dataset can be found here.
Open Datasets Yes Data can be downloaded here, including segmentation masks and better camera poses for existing datasets. ... We use multiple existing datasets and a new instructive synthetic dataset... D-Ne RF (Pumarola et al. (2020)) ... Nerfies (Park et al. (2021a)) and Hyper Ne RF (Park et al. (2021b)) ... Ne RF-DS (Yan et al. (2023)) ... The i Phone dataset from Dy Check (Gao et al. (2022)).
Dataset Splits Yes For all datasets, we use the original train/test splits as provided by the authors to ensure fair comparison.
Hardware Specification Yes All train time and rendering speeds are computed on NVIDIA Ge Force RTX 3090 cards.
Software Dependencies No The paper describes experimental implementations and hyperparameters, but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes For learning rates, people set spherical harmonics feature learning rate as 0.0025, opacity learning rate as 0.05, scaling learning rate as 0.005, rotation learning rate as 0.001, position learning rate as 0.00016 that exponentially decrease to 0.0000016 with learning rate delay 0.01, after 30000 iterations. Gaussian densification starts from 500th training iteration, and ends until 15000th training iteration. Densification gradient threshold is set to 0.0002. The spherical harmonics (SH) degree is set to be 3 in 3DGS, percent dense is set to be 0.01. ... Deformation Learning Rate starts from 0.00016, and exponentially decays to 0.0000016 with delay multiplier 0.01 after certain steps. ... network depth as 8; network width as 256; time embedding dim as 6 for synthetic, 10 for real-world; position embedding dim as 10...