Bundle Adjustment Makes or Breaks 3D Gaussian Splats, Study Finds

This study evaluates LEGS, a 3D language-guided mapping system, across large indoor environments using open-vocabulary object recall tasks. LEGS achieves comparable object localization performance to LERF but reduces training time from 44 minutes to 12 minutes. Experiments also show that global bundle adjustment significantly improves Gaussian splat reconstruction quality across camera configurations, reducing artifacts and enhancing photorealism—highlighting the critical role of camera setup and optimization in scalable 3D scene understanding.