Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Trick-GS: A Balanced Bag of Tricks for Efficient Gaussian Splatting

Anil Armagan, Albert Saà-Garriga, Bruno Manganelli, Mateusz Nowak, Mehmet Kerim Yucel

TL;DR

This work targets the practical bottlenecks of Gaussian splatting (GS) for 3D reconstruction on constrained devices. It introduces Trick-GS, a composite method that fuses progressive training, volume and significance-based pruning, SH masking, and accelerated training into a cohesive pipeline, achieving comparable accuracy to vanilla GS but with markedly reduced training time and storage. Across multiple benchmark datasets, Trick-GS demonstrates up to 2x faster training, up to 40x smaller disk size, and around 2x faster rendering, with a tunable design that can prioritize convergence, speed, accuracy, or storage. The approach is particularly relevant for on-device 3D reconstruction where memory and compute are limited, and it shows promise for further improvements via post-processing and automated device-aware adaptations.

Abstract

Gaussian splatting (GS) for 3D reconstruction has become quite popular due to their fast training, inference speeds and high quality reconstruction. However, GS-based reconstructions generally consist of millions of Gaussians, which makes them hard to use on computationally constrained devices such as smartphones. In this paper, we first propose a principled analysis of advances in efficient GS methods. Then, we propose Trick-GS, which is a careful combination of several strategies including (1) progressive training with resolution, noise and Gaussian scales, (2) learning to prune and mask primitives and SH bands by their significance, and (3) accelerated GS training framework. Trick-GS takes a large step towards resource-constrained GS, where faster run-time, smaller and faster-convergence of models is of paramount concern. Our results on three datasets show that Trick-GS achieves up to 2x faster training, 40x smaller disk size and 2x faster rendering speed compared to vanilla GS, while having comparable accuracy.

Trick-GS: A Balanced Bag of Tricks for Efficient Gaussian Splatting

TL;DR

This work targets the practical bottlenecks of Gaussian splatting (GS) for 3D reconstruction on constrained devices. It introduces Trick-GS, a composite method that fuses progressive training, volume and significance-based pruning, SH masking, and accelerated training into a cohesive pipeline, achieving comparable accuracy to vanilla GS but with markedly reduced training time and storage. Across multiple benchmark datasets, Trick-GS demonstrates up to 2x faster training, up to 40x smaller disk size, and around 2x faster rendering, with a tunable design that can prioritize convergence, speed, accuracy, or storage. The approach is particularly relevant for on-device 3D reconstruction where memory and compute are limited, and it shows promise for further improvements via post-processing and automated device-aware adaptations.

Abstract

Gaussian splatting (GS) for 3D reconstruction has become quite popular due to their fast training, inference speeds and high quality reconstruction. However, GS-based reconstructions generally consist of millions of Gaussians, which makes them hard to use on computationally constrained devices such as smartphones. In this paper, we first propose a principled analysis of advances in efficient GS methods. Then, we propose Trick-GS, which is a careful combination of several strategies including (1) progressive training with resolution, noise and Gaussian scales, (2) learning to prune and mask primitives and SH bands by their significance, and (3) accelerated GS training framework. Trick-GS takes a large step towards resource-constrained GS, where faster run-time, smaller and faster-convergence of models is of paramount concern. Our results on three datasets show that Trick-GS achieves up to 2x faster training, 40x smaller disk size and 2x faster rendering speed compared to vanilla GS, while having comparable accuracy.
Paper Structure (19 sections, 7 equations, 5 figures, 3 tables)

This paper contains 19 sections, 7 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Preliminaries
  5. Tricks for Learning Efficient 3DGS Representations
  6. Pruning with Volume Masking
  7. Pruning with Significance of Gaussians
  8. Spherical Harmonic (SH) Masking.
  9. Progressive Training.
  10. Progressive training by blurring.
  11. Progressive training by resolution.
  12. Progressive training by scales of Gaussians.
  13. Accelerated training for 3DGS
  14. Experimental Results
  15. Experimental Setup
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Qualitative comparison of the methods. Our method can recover more consistent text and background (top), or better structure metallic fence (bottom) while keeping low training time and great compression rates. We show zoomed prediction images (b-e) for the cropped areas with red rectangles as in ground-truth image (a) and green circles for improvements.
  • Figure 2: Number of Gaussians (#G) during training (on MipNeRF 360 - bicycle scene) for all methods, number of masked Gaussians (#Masked-G) and number of Gaussians with a masked SH band for our method. Our method performs a balanced reconstruction in terms of training efficiency by not letting the number of Gaussians increase drastically as other methods during training, which is a desirable property for end devices with low memory.
  • Figure 3: Impact of progressive training strategies on challenging background reconstructions. We empirically found that progressive training strategies as downsampling, adding Gaussian noise and changing the scale of learned Gaussians have a significant impact on the background objects with holes such as tree branches.
  • Figure 4: Visual results (a) from vanilla 3DGS (first row) and a model trained with progressive resolution based strategy (second row) starting with scale $0.125$. We use 'garden' from MipNeRF360 dataset and zoom into the improvements (b) & (c) for clarity.
  • Figure 5: PSNR and training time evaluations w.r.t the lowest scale used to start a progressive resolution-based training.