Trim 3D Gaussian Splatting for Accurate Geometry Representation


Lue Fan1,2 *     Yuxue Yang1 *     Minxing Li1     Hongsheng Li2,3      Zhaoxiang Zhang1
1CASIA     2MMLab, CUHK     3Shanghai AI Lab

TrimGS exhibits better geometric details and perceptual quality. In Figure (a) TrimGS separates the slender crossbars of a bench. Figure (b) showcases that TrimGS excels in rendering intricate details in both color and normals.

Abstract

In this paper, we introduce Trim 3D Gaussian Splatting (TrimGS) to reconstruct accurate 3D geometry from images. Previous arts for geometry reconstruction from 3D Gaussians mainly focus on exploring strong geometry regularization. Instead, from a fresh perspective, we propose to obtain accurate 3D geometry of a scene by Gaussian trimming, which selectively removes the inaccurate geometry while preserving accurate structures. To achieve this, we analyze the contributions of individual 3D Gaussians and propose a contribution-based trimming strategy to remove the redundant or inaccurate Gaussians. Furthermore, our experimental and theoretical analyses reveal that a relatively small Gaussian scale is a non-negligible factor in representing and optimizing the intricate details. Therefore the proposed TrimGS maintains relatively small Gaussian scales. In addition, TrimGS is also compatible with the effective geometry regularization strategies in previous arts. When combined with the original 3DGS and the state-of-the-art 2DGS, TrimGS consistently yields more accurate geometry and higher perceptual quality.

Reconstructions on Mip-NeRF 360 Dataset

Comparison of color and normal maps wtih 3DGS

Trim3DGS can reconstruct smoother normal maps than 3DGS.

garden

stump

Comparison of color and normal maps wtih 2DGS

Compared with 2DGS, our Trim2DGS exhibits better geometric details and perceptual quality.

bicycle

flowers

Comparison of test-set view rendering quality

Our Trim2DGS significantly enhances perceptual quality in high-frequency regions and substantially reduces the storage consumption.

Visualization of meshes

Visualization of Gaussian centers

Our Trim2DGS yields a more uniform distribution of Gaussians while reducing the storage consumption.

Reconstructions on DTU Dataset

Comparison of meshes with other methods

For better visualization, we zoom in the details, revealing that Trim2DGS and Trim3DGS exhibit accurate geometry representation.

Sample 1 (scan 24)

Sample 2 (scan 106)

Visualization of Gaussian centers

Our Trim2DGS and Trim3DGS yield a more uniform distribution of Gaussians, representing the geometry more accurately.

More samples

BibTeX

@article{fan2024trim,
  author    = {Fan, Lue and Yang, Yuxue and Li, Minxing and Li, Hongsheng and Zhang, Zhaoxiang},
  title     = {Trim 3D Gaussian Splatting for Accurate Geometry Representation},
  journal   = {arXiv preprint arXiv:2406.07499},
  year      = {2024},
}

Acknowledgements

We sincerely thank the great work 3D Gaussian Splatting and 2D Gaussian Splatting. Our Trim3DGS and Trim2DGS are built upon these foundations.