跨模态交互学习与迭代融合的3d视觉定位

Key Points

• The aim is to enhance 3D visual localization by addressing limitations in existing methods related to modality dependence and adaptability.
• Extracted features from point clouds and text using specific encoders.
• Implemented a transformer-based module for enhancing point cloud features.
• Utilized a symmetric interaction learning module to capture deep connections between point cloud and text features.
• Developed a cross-modal iterative fusion module for integrating information.
• Achieved accuracy scores of 86.19% and 69.68% in the unique subset of ScanRefer at Acc@0.25 and Acc@0.5.
• Attained a localization accuracy of 65.20% in the easy subset of Nr3D.
• Achieved a localization accuracy of 74.87% in the easy subset of Sr3D.
• Demonstrated consistent localization performance across multiple 3D scenes.