Event

PanoGraphGS unifies panoptic segmentation and hierarchical 3D scene graphs

1 reports · 1 sources

Jinghan Gao et al. propose the PanoGraphGS framework, which integrates panoptic 3D Gaussian representation learning and hierarchical 3D scene graph construction into a single framework, learning semantic and instance features on shared Gaussian geometry, and release NaviGS, a real-world dataset for panoptic scene understanding. The paper claims that the method outperforms the strongest comparison baseline by an average of more than 50 percentage points in object node accuracy on Replica and NaviGS, and achieves state-of-the-art performance in 2D panoptic segmentation among the evaluated methods.