[LichtFeld-Studio] ImprovedGS+: Redefining the Pareto Front for 3D Gaussian Splatting Efficiency

TL;DR

ImprovedGS+ is a high-efficiency reinvention of the 3D Gaussian Splatting (3DGS) pipeline. By abandoning high-level Python abstractions in favor of native C++/CUDA kernels, it achieves a significant performance leap: 26.8% faster training and 13.3% fewer Gaussians than SOTA baselines like MCMC, all while delivering superior visual quality (PSNR +1.28 dB over ADC).

The Bottleneck: Synchronization and Redundancy

Despite the explosive growth of 3DGS, current implementations suffer from two primary inefficiencies:

1. Computational Overhead: Moving data between Python/PyTorch and GPU kernels creates a synchronization bottleneck that throttles training speed.
2. Densification Drift: Many models place too many Gaussians in flat areas or fail to align primitives with actual geometric boundaries, leading to blurry edges or "stipple noise."

ImprovedGS+ addresses these by asking: How can we make densification mathematically deterministic and computationally native?

Methodology: The Native Advantage

1. Laplacian Filter & NMS for Geometry

To solve "densification drift," the authors implemented a distilled Canny Edge Detection logic directly in CUDA. By applying Non-Maximum Suppression (NMS), the model isolates fine geometric manifolds. This ensures that new Gaussians are only spawned on true structural backbones rather than noisy planar surfaces.

Figure: Comparison showing how ImprovedGS+ (c) produces a thinned structural backbone compared to the noisy baseline (b).

2. Long-Axis-Split (LAS) Kernel

Instead of standard cloning, ImprovedGS+ uses an exclusive Long-Axis-Split. The entire splitting logic—calculating offsets, updating logarithmic scales, and adjusting opacity—is encapsulated in a single CUDA kernel.

• Optimization Insight: By recognizing that displacement occurs only along the principal axis, they reduce the global coordinate transformation from 9 multiplications to just 3, significantly lowering the cost per split.

3. Two-Stage Learning Schedule

The method uses a strategic learning rate pivot:

• Stage I (Expansion): High-momentum scale learning (4x higher than standard) to fill the scene volume quickly.
• Stage II (Refinement): An Exponential Scale Scheduler decays the rate, "freezing" structural details and preventing primitives from oscillating or over-expanding.

Performance: Faster, Leaner, Sharper

The experimental results on the Mip-NeRF360 benchmark demonstrate a new state-of-the-art for resource-constrained reconstruction.

Key highlights include:

• Efficiency: The "1M Budget" variant saves ~17 minutes of training per session compared to MCMC.
• Outdoor Dominance: In the "Garden" scene, the method reached a PSNR of 29.39 dB, outperforming common baselines while maintaining a leaner Gaussian count.
• Ablation Success: Disabling the Exponential Scale Scheduler resulted in a 0.42 dB drop in PSNR for indoor scenes like "Bonsai," proving the necessity of the refinement phase.

Critical Analysis & Future Outlook

ImprovedGS+ succeeds by focusing on the "plumbing" of 3DGS. It proves that the path to real-time, high-fidelity 3D reconstruction isn't just about better loss functions, but about optimized GPU-native execution and deterministic growth strategies.

Limitations: Currently, the initial stages of the Laplacian pipeline still use multiple memory passes. The authors suggest that Kernel Fusion—merging blurring, gradient extraction, and NMS into a single monolithic kernel—could further reduce VRAM bandwidth in future iterations.

Takeaway: For researchers and engineers, ImprovedGS+ serves as a masterclass in how low-level hardware optimization (CUDA) and high-level structural intuition (NMS/LAS) can be combined to break established performance ceilings.