TL;DR
ACD-GS introduces a geometry-aware densification strategy for 3D Gaussian Splatting (3DGS) that replaces heuristic, gradient-based growth with a second-order depth curvature prior. By asymmetrically regulating Gaussian cloning and splitting, the framework slashes storage and point counts by nearly 60% while maintaining—and often improving—photorealistic rendering quality.
What Problem Does This Paper Address?
3D Gaussian Splatting has revolutionized real-time novel view synthesis, but it has a "bloat" problem. During training, the algorithm decides where to add more Gaussians (densification) based on first-order screen-space gradients.
The issue is that these gradients are geometry-blind. They respond to photometric signals, meaning a flat wall with a complex texture (like a brick pattern) might trigger the same massive proliferation of Gaussians as a complex geometric edge. This leads to redundant primitives, massive storage requirements, and "floaters"—artifact Gaussians that appear in empty space. Existing solutions often rely on post-hoc compression, but ACD-GS aims to prevent this redundancy at the source: during the training process itself.
Methods
The authors propose ACD-GS (Asymmetric Curvature-aware Densification), which moves away from purely photometric cues toward structural ones.
The framework consists of three core components:
- Curvature-Guided Asymmetric Densification: Instead of just looking at gradients, the model computes a "curvature proxy" using the discrete Laplacian of the rendered depth map. This second-order response is highly sensitive to actual geometric discontinuities (edges, thin structures) but remains quiet in planar, textured areas.
- Asymmetry: It decouples cloning and splitting. It suppresses cloning near edges to avoid "foreground-background duplication" (floaters) while promoting splitting to refine boundary-crossing Gaussians into finer pieces.
- Multi-view Photometric Error Gating: To ensure densification is truly necessary, the model checks candidates against eight auxiliary views. If an error is only visible from one view, it’s likely transient noise or an occlusion artifact and is ignored.
- Capacity-Constrained Adaptive Optimization: The model uses a scene-aware "growth quota" and an annealed opacity pruning strategy. This dynamically regulates the total number of Gaussians, ensuring the representation stays compact as training converges.
Datasets
The authors evaluated ACD-GS against several state-of-the-art baselines (including Mip-Splatting, AbsGS, and Pixel-GS) using three standard benchmarks:
- Mip-NeRF 360: Challenging indoor and outdoor scenes with complex backgrounds.
- Tanks & Temples: Large-scale outdoor environments.
- Deep Blending: Scenes with significant depth complexity and high-frequency textures.
Key Insights and Ideas
- Second-order is better for structure: First-order gradients capture change, but second-order depth responses (curvature) capture the change of change. This is the mathematical key to distinguishing a flat textured surface from a physical corner.
- Cloning vs. Splitting: Standard 3DGS treats these operations similarly. ACD-GS recognizes they have different geometric effects. Cloning adds capacity (good for smooth areas), while splitting refines boundaries (good for edges). By treating them asymmetrically based on curvature, the model places Gaussians exactly where they are needed.
- Native Sparsification: By curbing redundant growth during training, the model produces a compact representation "natively." This is more efficient than training a massive model and then trying to compress it later.
Key Results
The quantitative results are striking. Across all datasets, ACD-GS achieved:
- ~60% reduction in both Gaussian count and storage size compared to standard 3DGS and other recent baselines.
- Faster training times: Because there are fewer Gaussians to optimize and render during the training loop, ACD-GS often trains significantly faster than its competitors.
- Competitive Fidelity: Despite the massive reduction in points, rendering metrics (PSNR, SSIM) remained competitive or superior. On the Deep Blending dataset, ACD-GS achieved the highest average PSNR (25.12) among all tested methods.
Limitations and Future Work
While ACD-GS excels at native sparsification, the authors note that "extreme compression" can eventually degrade high-frequency texture fidelity due to the inherent capacity limits of explicit Gaussians.
To address this, they suggest future work could explore hybrid architectures, perhaps replacing explicit spherical harmonics with lightweight Multi-Layer Perceptrons (MLPs) to maintain detail under even stricter point budgets. Additionally, since ACD-GS is orthogonal to post-hoc compression (like quantization), combining the two could lead to even more ultra-compact models for low-bandwidth streaming.
Summary updated