Text2Loc++: Generalizing 3D Point Cloud Localization from Natural Language
Text2Loc++ is a novel neural network that aligns natural language descriptions with 3D point clouds for submap localization in a coarse-to-fine pipeline. It uses advanced techniques like Hierarchical Transformers and contrastive learning to create robust cross-modal embeddings. The system achieves up to 15% better accuracy on KITTI360Pose and introduces a new city-scale dataset.