Quantitative evaluation results on the ScanRefer dataset. Our CSVG method demonstrates superior performance compared to state-of-the-art zero-shot 3DVG methods, achieving significant improvements in accuracy metrics.
ScanRefer Dataset Results
Evaluation results on ScanRefer and Nr3D datasets
Quantitative evaluation results on the ScanRefer dataset. Our CSVG method demonstrates superior performance compared to state-of-the-art zero-shot 3DVG methods, achieving significant improvements in accuracy metrics.
Quantitative evaluation results on the Nr3D dataset. The results show consistent improvements across different evaluation metrics, demonstrating the effectiveness and robustness of our constraint satisfaction approach.
Qualitative comparison showing grounding results between our CSVG method and ZSVG3D. Our approach demonstrates more accurate object localization and better handling of complex spatial relationships in 3D scenes.