G$^2$ARD-GS: Geometry-Guided Anchor-Regularized Gaussian Splatting Distillation

2026-08-06

AI Summary

This work addresses the high storage, transmission, and rendering costs of 3D Gaussian Splatting models built from dense colored LiDAR data, which stem from an excessive number of primitives, while noting that aggressive compression often degrades geometric fidelity and harms novel view synthesis. To mitigate this, the authors propose a geometry-guided distillation approach that employs surface-aware progressive primitive clustering and optimizes appearance under anchor-based constraints within a fixed topology—without adding or removing any primitives. Evaluated on MatrixCity, the method achieves compression ratios of 5–30× with PSNR improvements of 3.2–6.8 dB over PUP. When reusing the frozen geometry for out-of-trajectory appearance adaptation, it further gains 3.7–4.9 dB in PSNR while maintaining accurate image-to-model registration on Cambridge King’s College.

Abstract

Dense colored LiDAR maps provide accurate city-scale geometry, but lifting them into 3D Gaussian Splatting (3DGS) retains millions of primitives, making the resulting models costly to store, transmit, render, and adapt. Aggressive primitive reduction alleviates this burden, but can remove the local surface support needed for stable novel-view synthesis and downstream geometric use. We introduce G$^2$ARD-GS, a geometry-guided distillation method that converts a dense Gaussian prior instantiated either as a training-free point-cloud lift or a trained GS model into a compact, reusable representation. G$^2$ARD-GS progressively consolidates the prior into surface-aware representatives, then recovers appearance on the resulting fixed topology under construction-time anchor constraints, with no primitives added or removed during recovery. Under limited supervision, geometry-aware view selection allocates the available view budget. On MatrixCity, G$^2$ARD-GS achieves the best PSNR, SSIM, and LPIPS across matched $5\times$--$30\times$ compression budgets, outperforming PUP by $3.2$--$6.8$,dB in PSNR. When reused as frozen geometry, the compact model improves off-trajectory appearance adaptation by $3.7$--$4.9$,dB over PUP 3D-GS and preserves image-to-model registration accuracy on Cambridge KingsCollege at $30\times$ compression. Project page: https://patrick1159.github.io/gardGS-page/.

Problem

• 3D Gaussian Splatting
• model compression
• geometry preservation
• novel-view synthesis
• LiDAR mapping

Innovation

• Geometry-Guided Distillation
• Gaussian Splatting
• Primitive Consolidation
• Anchor-Regularized Optimization
• Compact 3D Representation