SplatQ: Interaction-Aware Pruning of 3D Gaussian Splatting via Quadratic Selection

TL;DR

SplatQ introduces a combinatorial pruning method for 3D Gaussian Splatting that improves reconstruction quality and balance by modeling spatial and appearance-based block partitioning, smoothing budget allocation, and jointly optimizing Gaussian selection through a quadratic objective. It uses simulated annealing or a coherent Ising machine to solve the optimization, capturing interaction effects between Gaussians during deletion.

Why It Matters

This approach enables more efficient and visually consistent 3D scene reconstruction by preventing uneven quality distribution across regions, which is critical for real-time rendering and large-scale scene processing. The method enhances performance without sacrificing visual fidelity, making it suitable for applications in AR/VR and autonomous systems.

Key Findings

• SplatQ achieves better overall reconstruction quality compared to existing pruning methods by coordinating regional budget allocation and joint Gaussian selection.
• Budget smoothing prevents over-concentration of retained Gaussians in certain areas, improving local detail preservation and spatial balance.
• Modeling pairwise interactions through a quadratic approximation leads to higher-quality selections under fixed budget constraints.