3d-scene-graph

NoPA introduces a non-parametric distribution-based approach for real-time 3D scene graph generation, preserving geometric details using kernel density estimates and particle-based object representation, substantially outperforming current methods.

This paper proposes Transformation-Aware Decoupling (TAD), a framework for viewpoint-robust 3D scene graph generation that decouples relation reasoning into directional and invariant branches to handle yaw viewpoint changes, achieving state-of-the-art robustness on the 3DSSG benchmark.