P2P-Pose: point to point relative camera pose regression with image matching

Abstract Relative camera pose estimation is a fundamental task in computer vision. Existing learning-based methods often rely on global features for direct pose regression, which diminishes the geometric constraints and information interaction provided by local features, thereby limiting the accuracy of relative pose estimation. To address this limitation, we propose P2P-Pose, a point-to-point relative camera pose regression framework guided by image matching. P2P-Pose facilitates a comprehensive learning paradigm that regresses relative pose by simultaneously leveraging local matched point features and global point-level representations. Given an image pair, P2P-Pose employs a spatial attention interaction module to capture both intra- and inter-image correlations, explicitly constructs geometric constraints from local feature points through image matching, and integrates corresponding feature pairs using a channel attention fusion module. Subsequently, a dual-branch perceptron network regresses both local and global pose estimates, while an adaptive weighting strategy dynamically balances their respective contributions, effectively complementing local geometric information with global contextual understanding. Experimental results show that P2P-Pose achieves 86.8%/94.4% relative pose accuracy (for translation errors ⩽ 1 m and rotation errors ⩽ 30 ∘ ) on the Matterport3D dataset and 86.0%/84.8% on the ScanNet dataset, representing absolute improvements of up to 1.4%/0.2% and 4.0%/2.2% over state-of-the-art baselines. Evaluation on Cambridge Landmarks further confirms its competitive performance and robustness in outdoor environments.