Researchers have developed a new framework called ExtrinSplat for open-vocabulary understanding in 3D scenes. The framework decouples geometry from semantics by clustering Gaussians into object groups and using a vision-language model to generate textual hypotheses. This approach reduces scene adaptation time and storage overhead compared to traditional embedding-based methods. --- Why it matters: This matters because it addresses the limitations of current methods for open-vocabulary understanding in 3D scenes, which can be slow and require large amounts of memory. ExtrinSplat's efficiency and effectiveness could lead to improved performance in applications such as robotics and autonomous vehicles. Source: https://arxiv.org/abs/2509.22225