Reproduce “LEGS: Laplacian-Enhanced Gaussian Splatting with a Nonlinear Weighted Loss” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongfei Guo, Qizhou Huo, Xuan Sun, Yuanhao Gong
LEGS adds structural pixel weighting to Gaussian Splatting scene reconstruction from photographs, without changing its representation or renderer. It computes normalized Laplacian responses from input images and maps them to weights in the photometric loss. On training views, it reports Mip-NeRF360 PSNR of 31.0006, versus 29.3242 for 3DGS and 30.4814 for EGGS; this does not establish unseen-view performance.
Key insights
- Second-order Laplacian operators provide a more comprehensive structural cue than first-order gradients, effectively capturing fine details and complex scene geometry for 3DGS optimization.
- Nonlinear weighting functions for structural responses offer superior flexibility and effectiveness compared to linear mappings, allowing for nuanced control over the contribution of different structural levels to the loss.
- Enhancing structure awareness in radiance field reconstruction can be achieved efficiently and modularly at the loss level, preserving the inherent advantages of the 3DGS framework while improving reconstruction fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.