Learning Overlap Detection for Domain-Adaptive Image-to-Point Cloud Registration

Zhixin Cheng, Bohao Liao, Xinjun Li, Baoqun Yin, Wenfei Yang, Tianzhu Zhang

Keywords: Intrinsic Geometric, Domain Adaption, I2P

Abstract:

Outdoor registration methods often employ an dedicated module to detect overlapping regions between images and point clouds.
While effective, this strategy is not directly applicable to indoor scenarios and increases computational cost. However, to further improve indoor registration accuracy, it is crucial to identify and isolate overlapping regions, minimizing interference from non-overlapping areas.
Furthermore, without targeted design, aligning image and point cloud features may lead to mismatches during feature interaction.
To address these issues, we propose two modules: the Reinforcement Learning Overlap Detector (RLOD) and the Hierarchical Domain Adaptation Interaction (HDAI) module. RLOD adaptively selects overlapping regions by leveraging intrinsic geometric information, thus constraining the matching space and improving accuracy. HDAI aligns image and point cloud features at both mean and covariance levels, mitigating cross-modal discrepancies and stabilizing attention. Experiments on RGB-D Scenes v2 and 7-Scenes benchmarks demonstrate that our method achieves superior performance, setting a new state of the art for image-to-point cloud registration.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 1475