Re-Activating Frozen Primitives for 3D Gaussian Splatting

TL;DR

ReAct-GS tackles over-reconstruction in 3D Gaussian Splatting by identifying gradient magnitude dilution and a primitive frozen phenomenon as core causes. It introduces importance-aware densification to re-activate stalled growth and an adaptive re-activation mechanism comprising density-guided cloning and needle-shape perturbation to revive frozen primitives. Across 13 real-world scenes, ReAct-GS yields state-of-the-art performance on standard novel view synthesis metrics and preserves geometric details, while remaining parameter-efficient and transferable to other 3D-GS variants such as Pixel-GS. The work advances 3D-GS by addressing fundamental optimization bottlenecks, enabling more reliable reconstruction in complex scenes and broader applicability.

Abstract

3D Gaussian Splatting (3D-GS) achieves real-time photorealistic novel view synthesis, yet struggles with complex scenes due to over-reconstruction artifacts, manifesting as local blurring and needle-shape distortions. While recent approaches attribute these issues to insufficient splitting of large-scale Gaussians, we identify two fundamental limitations: gradient magnitude dilution during densification and the primitive frozen phenomenon, where essential Gaussian densification is inhibited in complex regions while suboptimally scaled Gaussians become trapped in local optima. To address these challenges, we introduce ReAct-GS, a method founded on the principle of re-activation. Our approach features: (1) an importance-aware densification criterion incorporating $α$-blending weights from multiple viewpoints to re-activate stalled primitive growth in complex regions, and (2) a re-activation mechanism that revitalizes frozen primitives through adaptive parameter perturbations. Comprehensive experiments across diverse real-world datasets demonstrate that ReAct-GS effectively eliminates over-reconstruction artifacts and achieves state-of-the-art performance on standard novel view synthesis metrics while preserving intricate geometric details. Additionally, our re-activation mechanism yields consistent improvements when integrated with other 3D-GS variants such as Pixel-GS, demonstrating its broad applicability.

Figures (9)

• Figure 1: Visualization of renderings, depth maps and point cloud distributions in an unbounded 3D scene. The method Abs-GS ye2024absgs, relying on large-scale splitting strategy (upper), leads to ill-position Gaussian growth in complex regions.
• Figure 2: Illustration of primitive frozen phenomenon and our adaptive re-activation mechanism. (a) and (b) depict the underlying factors causing small-scale and needle-shape 3D Gaussians to freeze; (c) and (d) showcase our proposed re-activation strategies for these respective frozen primitives.
• Figure 3: Visualization of primitive frozen phenomenon across optimization stages. Top: Small-scale 3D Gaussians freeze locally and barely exhibit growth to relieve blurry during densification; Bottom: Needle-shape primitives persist and cannot be mitigated in post-densification optimization.
• Figure 4: Qualitative comparisons of different methods on scenes from the Mip-NeRF 360 and Tanks & Temples datasets. Close-up views highlight the challenging areas with high-frequency details, where over-reconstruction is particularly pronounced.
• Figure 5: Comparison of depth map renderings across different methods on indoor/outdoor scenes from the Mip-NeRF 360 dataset. Enlarged views highlight the challenging areas where geometric consistency and coherence are crucial.