2DGS-R: Revisiting the Normal Consistency Regularization in 2D Gaussian Splatting

TL;DR

2DGS-R addresses the enduring conflict between rendering fidelity and geometric accuracy in 2D Gaussian Splatting by introducing a hierarchical, three-stage training regime. It starts with geometry-focused NC-enabled initialization, followed by error-driven refinement of appearance for high-error Gaussians, and culminates in targeted enhancement via in-place cloning with opacity frozen to preserve structure. The approach yields high-quality renderings while maintaining fine geometric reconstruction with only ~1% extra storage and minimal training overhead, validated across DTU, Tanks & Temples, and NeRF-Synthetic datasets. This makes 2DGS-R a practical and scalable improvement for neural scene representations based on 2D Gaussian primitives.

Abstract

Recent advancements in 3D Gaussian Splatting (3DGS) have greatly influenced neural fields, as it enables high-fidelity rendering with impressive visual quality. However, 3DGS has difficulty accurately representing surfaces. In contrast, 2DGS transforms the 3D volume into a collection of 2D planar Gaussian disks. Despite advancements in geometric fidelity, rendering quality remains compromised, highlighting the challenge of achieving both high-quality rendering and precise geometric structures. This indicates that optimizing both geometric and rendering quality in a single training stage is currently unfeasible. To overcome this limitation, we present 2DGS-R, a new method that uses a hierarchical training approach to improve rendering quality while maintaining geometric accuracy. 2DGS-R first trains the original 2D Gaussians with the normal consistency regularization. Then 2DGS-R selects the 2D Gaussians with inadequate rendering quality and applies a novel in-place cloning operation to enhance the 2D Gaussians. Finally, we fine-tune the 2DGS-R model with opacity frozen. Experimental results show that compared to the original 2DGS, our method requires only 1\% more storage and minimal additional training time. Despite this negligible overhead, it achieves high-quality rendering results while preserving fine geometric structures. These findings indicate that our approach effectively balances efficiency with performance, leading to improvements in both visual fidelity and geometric reconstruction accuracy.

Figures (10)

• Figure 1: (a) Rendered color. (b) Rendered normal. (c) Normal from rendered depth. Our motivation for this work is to achieve high-quality rendering and precise surface reconstruction for 2DGS huang20242dgs at the same time. Normal consistency regularization (NC) means minimizing the difference between (b) and (c).
• Figure 2: Each scene's average value of 2DGS on the Tanks and Temples Dataset Knapitsch2017. The introduction of NC increases $K_{a}$ and opacity. The larger the $K_a$ and opacity, the better the reconstruction quality, while the rendering quality will deteriorate.
• Figure 3: Our training process is divided into three stages. To illustrate this, we use a pure blue triangle as the ground truth example. Each ellipse denotes a 2D Gaussian, and the color variations reflect the evolution of appearance as training progresses.
• Figure 4: $\mu$ and $\Sigma$ represent the geometric properties of the Gaussians, while $\alpha$ and SHs are used to represent the appearance model of the Gaussians.
• Figure 5: We take a single Gaussian and two perspectives as examples to demonstrate how to calculate $E_{i}$.