New Method Improves 3D Scene Graph Robustness to Viewpoint Changes

A new framework called Transformation-Aware Decoupling (TAD) has been introduced to enhance the robustness of 3D Scene Graph Generation (3DSGG) models, particularly when dealing with varying viewpoints. Traditional 3DSGG models often struggle to maintain consistent relation predictions when an agent's yaw rotation changes, leading to an empirical mismatch in how different types of predicates should behave. TAD addresses this by decoupling relation reasoning into two distinct parts: one that learns cues for predicates that should remain stable regardless of viewpoint (like "attached to"), and another for directional predicates (like "left" or "front") that should transform with the observation frame. By merging these two branches and employing transformation-specific descriptors, TAD achieves state-of-the-art robustness to yaw viewpoint changes on the 3DSSG benchmark, without needing rotation augmentation during training, while maintaining competitive overall performance.