CA-I2P: Channel-Adaptive Registration Network with Global Optimal Selection

2025-06-26

AI Summary

In image-to-point-cloud registration without detection, performance degradation arises from inter-modal channel attention discrepancies and spurious correspondences induced by structural scene similarity. To address these challenges, this paper proposes an end-to-end robust registration framework. Its core contributions are: (1) a Channel-Adaptive Adjustment (CAA) module that dynamically calibrates cross-modal feature channel sensitivity; and (2) a Global Optimal Selection (GOS) module that replaces local nearest-neighbor matching with global optimization to suppress erroneous correspondences. The framework integrates a coarse-to-fine strategy, channel-attention-guided matching, and feature decoupling for enhanced representation learning—specifically tailored to heterogeneous RGB-D images and sparse point clouds. Evaluated on RGB-D Scenes V2 and 7-Scenes, the method achieves state-of-the-art performance, significantly improving both dense pixel-to-point correspondence accuracy and registration stability.

Technology Category

• Computer Vision: Multi-modal Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Multimodal Learning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Systems and Infrastructure for Web, Mobile and WoT: Experiences and lessons learnt from Web-based algorithms and system deployments
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Detection-free methods typically follow a coarse-to-fine pipeline, extracting image and point cloud features for patch-level matching and refining dense pixel-to-point correspondences. However, differences in feature channel attention between images and point clouds may lead to degraded matching results, ultimately impairing registration accuracy. Furthermore, similar structures in the scene could lead to redundant correspondences in cross-modal matching. To address these issues, we propose Channel Adaptive Adjustment Module (CAA) and Global Optimal Selection Module (GOS). CAA enhances intra-modal features and suppresses cross-modal sensitivity, while GOS replaces local selection with global optimization. Experiments on RGB-D Scenes V2 and 7-Scenes demonstrate the superiority of our method, achieving state-of-the-art performance in image-to-point cloud registration.

Problem

• Addresses degraded matching in image-point cloud registration
• Reduces redundant cross-modal correspondences from similar structures
• Improves registration accuracy via channel-adaptive and global optimization

Innovation

• Channel Adaptive Adjustment Module enhances features
• Global Optimal Selection replaces local selection
• Achieves state-of-the-art registration performance