Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LGDWT-GS: Local and Global Discrete Wavelet-Regularized 3D Gaussian Splatting for Sparse-View Scene Reconstruction

Shima Salehi, Atharva Agashe, Andrew J. McFarland, Joshua Peeples

TL;DR

LGDWT-GS tackles the problem of unreliable geometry and over-smoothed details in few-shot 3D reconstruction by embedding dual-frequency regularization via global and patch-wise Discrete Wavelet Transform losses into 3D Gaussian Splatting. The method preserves large-scale structure while recover­ing fine textures, and extends to multispectral RGB+NIR with shared geometry and cross-spectral supervision. A new four-band multispectral greenhouse dataset and an open-source few-shot benchmarking package are released to standardize evaluation. Across standard benchmarks and the greenhouse dataset, LGDWT-GS demonstrates sharper, more stable, and spectrally coherent reconstructions than strong baselines, highlighting the utility of frequency-aware priors in data-limited settings.

Abstract

We propose a new method for few-shot 3D reconstruction that integrates global and local frequency regularization to stabilize geometry and preserve fine details under sparse-view conditions, addressing a key limitation of existing 3D Gaussian Splatting (3DGS) models. We also introduce a new multispectral greenhouse dataset containing four spectral bands captured from diverse plant species under controlled conditions. Alongside the dataset, we release an open-source benchmarking package that defines standardized few-shot reconstruction protocols for evaluating 3DGS-based methods. Experiments on our multispectral dataset, as well as standard benchmarks, demonstrate that the proposed method achieves sharper, more stable, and spectrally consistent reconstructions than existing baselines. The dataset and code for this work are publicly available

LGDWT-GS: Local and Global Discrete Wavelet-Regularized 3D Gaussian Splatting for Sparse-View Scene Reconstruction

TL;DR

LGDWT-GS tackles the problem of unreliable geometry and over-smoothed details in few-shot 3D reconstruction by embedding dual-frequency regularization via global and patch-wise Discrete Wavelet Transform losses into 3D Gaussian Splatting. The method preserves large-scale structure while recover­ing fine textures, and extends to multispectral RGB+NIR with shared geometry and cross-spectral supervision. A new four-band multispectral greenhouse dataset and an open-source few-shot benchmarking package are released to standardize evaluation. Across standard benchmarks and the greenhouse dataset, LGDWT-GS demonstrates sharper, more stable, and spectrally coherent reconstructions than strong baselines, highlighting the utility of frequency-aware priors in data-limited settings.

Abstract

We propose a new method for few-shot 3D reconstruction that integrates global and local frequency regularization to stabilize geometry and preserve fine details under sparse-view conditions, addressing a key limitation of existing 3D Gaussian Splatting (3DGS) models. We also introduce a new multispectral greenhouse dataset containing four spectral bands captured from diverse plant species under controlled conditions. Alongside the dataset, we release an open-source benchmarking package that defines standardized few-shot reconstruction protocols for evaluating 3DGS-based methods. Experiments on our multispectral dataset, as well as standard benchmarks, demonstrate that the proposed method achieves sharper, more stable, and spectrally consistent reconstructions than existing baselines. The dataset and code for this work are publicly available
Paper Structure (18 sections, 6 equations, 9 figures, 4 tables)

This paper contains 18 sections, 6 equations, 9 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Few-Shot Novel View Synthesis
  4. 3DGS and Sparse-View Extensions
  5. Frequency-Aware Supervision
  6. Multispectral and Agricultural 3D Reconstruction
  7. Method
  8. LGDWT-GS Framework
  9. Global DWT Supervision
  10. Patch-wise DWT Supervision
  11. MultiSpectral Extension
  12. Dataset
  13. Experiments
  14. Benchmarking
  15. Quantitative Results
  16. ...and 3 more sections

Figures (9)

  • Figure 1: Overview of the LGDWT-GS framework. The model introduces frequency-domain regularization through global and local DWT losses. Combined supervision (L1, SSIM, and global DWT and local DWT) enhances structural stability and textural fidelity. Black and blue arrows represent operation and gradient flows, respectively.
  • Figure 2: Wavelet decomposition of the input image into four subbands: (a) LL (approximation), (b) LH (horizontal), (c) HL (vertical), (d) HH (diagonal).
  • Figure 3: $E_{LF}$ map used for patch selection: (a) Ground Truth, (b) $E_{LF}$ Map. Red regions denote low $E_{LF}$ values, indicating weak LF stability or missing HF details and revealing spatial frequency imbalance in the reconstruction.
  • Figure 4: Multispectral LGDWT-3DGS framework. Pseudo-RGB images (constructed from Red, Green, and Red-Edge bands) are used for COLMAP-based pose estimation and sparse reconstruction. RGB and NIR chanells are then jointly optimized under a shared geometry using cross-spectral supervision and DWT-based frequency regularization, improving geometric consistency and spectral alignment under sparse-view scenarios.
  • Figure 5: Example spectral channels for three representative plant scenes. Columns correspond to 580 nm (Green), 660 nm (Red), 735 nm (Red Edge), and 820 nm (NIR) bands. The final column shows the pseudo-RGB composite used for COLMAP reconstruction.
  • ...and 4 more figures