3D Scene Assertion Verification

Existing 3D Visual Question Answering (3D-VQA) methods rely on generative outputs that can be ambiguous in decision-making settings. We introduce 3D Scene Assertion Verification, a task that verifies natural language assertions in 3D scenes with strict binary judgments. We present 3DSAV, a large-scale diagnostic benchmark with 22.5k samples across six semantic types. To address this task, we propose DualLPSS, which uses dual-stage subspace routing for type-aware cross-modal fusion and scene-guided assertion focusing. Experiments show that DualLPSS achieves state-of-the-art performance on 3DSAV and handles complex logical assertions better than existing 3D-VQA baselines.

Many AI systems can describe a 3D scene, but they often do not give the clear yes-or-no answers needed by robots, AR systems, or other applications that must act on a scene. This paper studies how to check whether a statement about a 3D scene is true or false. We build 3DSAV, a dataset of 22,520 verified statements about indoor scenes, and introduce DualLPSS, a model designed for this verification task. In experiments, DualLPSS is more accurate than existing 3D question-answering models, especially on statements with harder logic.