NoPA introduces a non-parametric representation for online 3D scene graph generation. Instead of compressing each object into a single 3D Gaussian, it maintains a fixed particle set per object to preserve richer geometric structure while retaining real-time inference. The method uses maximum mean discrepancy on kernel density estimates for more robust candidate merging during online exploration, and propagates relations between highly affiliated objects to reduce relation loss caused by object misclassification. The paper reports substantial improvements over existing methods without sacrificing real-time speed.
No heat snapshots are available in the last 24 hours.