LEGS Trains 3.5x Faster Than LERF in Large-Scale Indoor Mapping

The paper introduces Language-Embedded Gaussian Splats (LEGS), a system that integrates CLIP embeddings into 3D Gaussian splatting for large indoor scenes, achieving 3.5x faster training than LERF while addressing pose drift through incremental bundle adjustment—though challenges remain in dynamic scenes, object ambiguity, and large-scale querying.