SDF-Prop FastGS: Structure-Driven Densification and Pruning for Textureless and Large-Scale Scenes

TL;DR: Ask a simple geometry brain (SDF) where surfaces are, then grow or trim Gaussians along those surfaces so we don’t waste time on empty space. We fuse FastGS’s multi-view consistency with an SDF-guided propagation and co-regularized training to handle low-texture, large-scale areas. First experiment: on ScanNet/ScanNet++ and Waymo-scale scenes, test if SDF-guided densification reduces Gaussians and training steps while improving mesh and image quality.

Research Question: Can a lightweight SDF and progressive propagation strategy, modulated by multi-view consistency, reduce the number of training iterations and Gaussians needed in textureless or unbounded scenes without harming photometric fidelity?

Hypothesis: Joint optimization with an SDF (for surface continuity) and consistency-aware propagation will place Gaussians more accurately, enabling stronger early pruning and fewer densification cycles, yielding faster convergence and better geometry in textureless regions.

Experiment Plan: - Setup: Add a small neural SDF trained online; steer densification along high |∇SDF| with normals alignment; use GaussianPro-like patch-guided propagation to seed new Gaussians with good orientation; add cross-view assisted training to enforce structural coherence; prune by low SDF-support and high cross-view inconsistency.

• Data/Materials: ScanNet/ScanNet++, Mip-NeRF 360, and large-scale outdoor datasets; evaluate with identical training budgets to FastGS.
• Measurements: PSNR/SSIM/LPIPS; Chamfer-L1 and normal consistency of extracted surfaces; #Gaussians and training steps to reach a target PSNR; performance in low-texture areas.
• Expected Outcomes: Faster convergence than vanilla FastGS, fewer Gaussians, improved geometry on textureless walls/floors, competitive or better photometric metrics.

References: ['Ren, S., Wen, T., Fang, Y., & Lu, B. (2025). FastGS: Training 3D Gaussian Splatting in 100 Seconds.', 'Xiang, H., Li, X., Lai, X., Zhang, W., Liao, Z., Cheng, K., & Liu, X.-P. (2024). GaussianRoom: Improving 3D Gaussian Splatting with SDF Guidance and Monocular Cues for Indoor Scene Reconstruction. IEEE International Conference on Robotics and Automation.', 'Cheng, K., Long, X., Yang, K., Yao, Y., Yin, W., Ma, Y., Wang, W., & Chen, X. (2024). GaussianPro: 3D Gaussian Splatting with Progressive Propagation. International Conference on Machine Learning.', 'Xiao, H., Zou, J., Zhou, Y., He, Y., & Kang, W. (2025). SplatCo: Structure-View Collaborative Gaussian Splatting for Detail-Preserving Rendering of Large-Scale Unbounded Scenes. arXiv.org.']