MGPose: Wide-Baseline Relative Camera Pose Estimation Using Matching-Guided Dual Channel-Attention

Abstract

Relative camera pose estimation is a fundamental task in computer vision and robotics. In wide- baseline scenarios with limited visual overlap, traditional methods often perform poorly. Existing deep learning approaches are also hindered by irrelevant features and insufficient modeling of the relative motion between image pairs, making accurate pose estimation particularly challenging. In this paper, we propose MGPose, a camera relative pose estimation method using a matching-guided dual-channel attention mechanism. For wide- baseline image pairs, MGPose effectively reduces interference from uncorrelated features through a feature matching strategy, utilizes camera motion prior knowledge to capture the relative motion characteristics of matched points, and employs a bidirectional channel cross-attention mechanism along with a channel self-attention mechanism to fully capture the interactions between different channels of matched points, enabling efficient feature fusion for the image pairs. Extensive experiments on Matterport3D and ScanNet show that MGPose outperforms or matches state-of-the-art methods in camera relative pose estimation.