First author · Technical white paper · UC Santa Barbara · June 2026

This method tracks how Gaussians expand into different planes from geometrically consistent MASt3R-COLMAP root points. The central idea is that the object of interest appears across many photos and remains geometrically consistent, while unwanted regions such as sky and peripheral noise do not map back to the same stable roots. Low-confidence branches can therefore be pruned with their descendants.

My contribution

• Engineered a lineage-based Gaussian filtering pipeline for high-fidelity building isolation without segmentation masks and documented the methodology in a technical white paper.
• Developed a custom PyTorch optimization pipeline for 3D Gaussian Splatting using SAM-derived semantic labels, and built a ReAct agent system to identify scene objects and remove object-level Gaussian artifacts.

White paper PDF Open interactive showcase

Move around the interactive showcase with WASD. The reconstructions come from student-taken photos and use this method to clean and isolate buildings, so they can be explored from viewpoints outside the original capture path.