CEM-4DGS: Clustered Error Correction with Grouped 4D Gaussian Splatting

Abstract

Existing 4D Gaussian Splatting (4DGS) methods struggle to accurately reconstruct dynamic scenes, often failing to resolve ambiguous pixel correspondences and inadequate densification in dynamic regions. We address these issues by introducing a novel method composed of two key components: (1) Elliptical Error Clustering and Error Correcting Splat Addition that pinpoints dynamic areas to improve and initialize fitting splats, and (2) Grouped 4D Gaussian Splatting that improves consistency of mapping between splats and represented dynamic objects.

Specifically, we classify rendering errors into missing-color and occlusion types, then apply targeted corrections via backprojection or foreground splitting guided by cross-view color consistency. Evaluations on Neural 3D Video and Technicolor datasets demonstrate state-of-the-art perceptual rendering quality, improving 0.39 dB PSNR on the Technicolor Light Field dataset, with better temporal consistency and object–splat alignment.

Method

Dynamic 4DGS often fails in two places: Gaussians do not stay attached to moving objects, and densification misses thin or fast motion. CEM-4DGS attacks both with a representation step and a correction step.

Grouped 4D Gaussians split groups when motion within a group diverges, improving object–splat consistency. Residual errors are clustered into elliptical regions and classified as missing-color or occlusion; new splats are added by backprojection or foreground split using cross-view color cues.

Interactive 3D Viewer

Explore the reconstructed 4D sequence interactively. Drag to orbit, scroll/pinch to zoom, use arrow keys or WASD to navigate, and play or scrub the timeline in the viewer panel. The viewer runs entirely in real-time WebGL using our exported Gaussian splat weights.

Results

On Technicolor and Neural 3D Video, CEM-4DGS recovers finer dynamic detail (fabric, fur, thin structures) and more coherent free-view motion than recent 4DGS baselines, with a +0.39 dB PSNR gain on Technicolor Light Field.

Video Results

GT / baseline / ours on hard motion.

Side-by-side vs. recent 4DGS methods.

Quantitative Results

Perceptual quality on Technicolor Light Field and Neural 3D Video. CEM-4DGS achieves the best PSNR and competitive LPIPS / DSSIM among recent 4DGS methods.

Technicolor Light Field

Method	PSNR ↑	DSSIM1 ↓	DSSIM2 ↓	LPIPS ↓
DyNeRF	31.80	—	0.021	0.140
HyperReel	32.73	0.047	—	0.109
4DGS	29.54	0.065	0.032	0.149
4DGaussians	30.79	0.079	0.040	0.178
STG	33.56	0.040	0.019	0.084
SWinGS	33.65	0.033	—	0.117
E-D3DGS	33.24	0.047	—	0.100
Ex4DGS	33.62	0.042	0.019	0.088
CEM-4DGS (Ours)	34.04	0.040	0.018	0.081

Neural 3D Video (selected methods)

Method	PSNR ↑	DSSIM1 ↓	DSSIM2 ↓	LPIPS ↓
4DGS	32.01	—	0.014	0.055
4DGaussians	31.15	—	0.016	0.049
STG	32.05	0.026	0.014	0.044
E-D3DGS	31.31	0.028	—	0.037
Ex4DGS	32.11	0.030	0.015	0.048
CEM-4DGS (Ours)	32.23	0.028	0.015	0.047

BibTeX

@inproceedings{kang2025cem4dgs,
  title        = {Clustered Error Correction with Grouped 4D Gaussian Splatting},
  author       = {Kang, Taeho and Park, Jaeyeon and Lee, Kyungjin and Lee, Youngki},
  booktitle    = {SIGGRAPH Asia 2025 Conference Papers},
  month        = {December},
  year         = {2025},
  pages        = {1--12},
  publisher    = {Association for Computing Machinery},
  address      = {New York, NY, USA},
  doi          = {10.1145/3757377.3763858},
  url          = {https://doi.org/10.1145/3757377.3763858}
}