Answers are grounded in this paper and cite their sources.
3D visual grounding aims to localize target objects in point-cloud scenes based on natural-language descriptions. However, incomplete geometry can hinder accurate object-center estimation, insufficient semantic cues can make target discrimination difficult, and similar objects can introduce relational ambiguity. To address these challenges, we propose URTOME, an Uncertainty- and Relation-Aware Transformer that integrates complementary geometric and linguistic modeling. Uncertainty-Aware Voting estimates point-wise voting uncertainty to refine seed-to-center offsets. Multi-Granularity Text Context Enhancement enriches language representations through implicit semantic enrichment and explicit description diversification. The Relation-Aware Transformer models directional and proximity relations among candidate objects to facilitate cross-modal matching. On ScanRefer, URTOME achieves 49.38% Acc@0.25 and 37.32% Acc@0.5, outperforming the 3DVG-Transformer baseline by 4.23 and 4.77 percentage points, respectively. It also achieves overall accuracies of 53.6% and 64.1% on Nr3D and Sr3D, respectively. Ablation studies demonstrate the complementary contributions of the three components, while proposal-level evaluation shows that uncertainty-aware voting improves proposal coverage and reduces vote-center error. These results highlight the effectiveness of integrating voting uncertainty, enriched textual context, and spatial relations for accurate 3D visual grounding.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: