Opacity-Gradient Driven Density Control for Compact and Efficient Few-Shot 3D Gaussian Splatting

TL;DR

This work addresses the inefficiency of few-shot 3D Gaussian Splatting by replacing the traditional view-space positional gradient densification with an error-driven trigger based on the opacity gradient, and by pairing it with a conservative multi-stage pruning strategy. The resulting framework produces dramatically more compact representations (over 40% fewer primitives on LLFF and ~70% on Mip-NeRF 360) while maintaining competitive reconstruction quality, achieving a new state-of-the-art on the efficiency-quality Pareto frontier for few-shot view synthesis. A depth-correlation loss provides geometric guidance in sparse data settings, and the combination yields fast rendering (high FPS) and reduced memory footprint. While effective, the approach remains dependent on external geometric priors and involves hyperparameters that may require tuning for different datasets, highlighting avenues for future end-to-end integration and adaptive pruning strategies.

Abstract

3D Gaussian Splatting (3DGS) struggles in few-shot scenarios, where its standard adaptive density control (ADC) can lead to overfitting and bloated reconstructions. While state-of-the-art methods like FSGS improve quality, they often do so by significantly increasing the primitive count. This paper presents a framework that revises the core 3DGS optimization to prioritize efficiency. We replace the standard positional gradient heuristic with a novel densification trigger that uses the opacity gradient as a lightweight proxy for rendering error. We find this aggressive densification is only effective when paired with a more conservative pruning schedule, which prevents destructive optimization cycles. Combined with a standard depth-correlation loss for geometric guidance, our framework demonstrates a fundamental improvement in efficiency. On the 3-view LLFF dataset, our model is over 40% more compact (32k vs. 57k primitives) than FSGS, and on the Mip-NeRF 360 dataset, it achieves a reduction of approximately 70%. This dramatic gain in compactness is achieved with a modest trade-off in reconstruction metrics, establishing a new state-of-the-art on the quality-vs-efficiency Pareto frontier for few-shot view synthesis.

Figures (4)

• Figure 1: An overview of the proposed optimization framework. Starting from a sparse set of input views and an initial point cloud from SfM, our method iteratively refines a set of 3D Gaussians. The core optimization loop consists of rendering, loss computation (photometric and geometric), and our revised Adaptive Density Control (ADC). Crucially, densification is driven by the opacity gradient, a direct proxy for rendering error, while a multi-stage pruning strategy removes transparent or redundant primitives. This cycle of rendering, error-driven adaptation, and pruning produces an efficient scene representation.
• Figure 2: Best-case qualitative results. Our framework produces high-fidelity reconstructions on diverse scenes while maintaining a highly compact model, demonstrating the effectiveness of the efficiency-focused optimization cycle.
• Figure 3: Marginal-case qualitative results. In scenes with challenging textures or materials, our framework makes a direct trade-off, accepting minor losses in high-frequency detail for significant gains in model compactness and rendering efficiency.
• Figure 4: Poor-performance (failure case) results. These examples illustrate the framework's limitations. When faced with highly ambiguous geometry or poor initial priors from SfM, our method can produce artifacts such as floaters or geometric inaccuracies.