PrismVG: Posterior State Fusion and Slot Evidence Verification for monocular multi-object 3D visual grounding

Monocular multi-object 3D visual grounding aims to localize all 3D targets that satisfy a natural language description within a single RGB image. Existing methods often rely on shallow fusion of 2D appearance and 3D geometric states, while holistic text–target matching tends to entangle heterogeneous semantics, thereby limiting representation completeness and causing partial-match bias. To address these issues, this paper proposes a two-stage framework, PrismVG. Specifically, we design Posterior State Fusion (PSF), which models the 3D geometric state as the prior of the object state and the 2D cropped image as the observation, and constructs a more complete and consistent object representation through confidence-aware posterior fusion. Meanwhile, we propose Slot Evidence Verification (SEV), which explicitly decomposes textual semantics into four attribute slots, namely appearance, geometry, visibility, and spatial information, and achieves finer-grained text–target alignment through attribute-wise evidence retrieval and consistency verification. Experiments on two benchmark datasets, MonoMulti3D-ROPE (44,342 images, 86,674 multi-object descriptions, and 535,149 object instances) and Mono3DRefer (2025 images and 41,140 referring expressions), demonstrate the effectiveness of PrismVG. On MonoMulti3D-ROPE, PrismVG improves the F1 score over the strongest baseline by 2.52 points under GT bboxes and 4.59 points under Pred bboxes. On Mono3DRefer, it further achieves gains of 3.57 and 4.13 F1 points under the two settings, respectively. These results demonstrate that PrismVG improves candidate representation learning and attribute-level text–object matching under both ideal and detector-predicted candidate settings.