EIMPNet: End-to-End Iterative Feature Matching and Pose Estimation Joint Optimization Network

Traditional pose estimation methods typically involve two separate steps: first using feature matching algorithms to identify matches and then estimating the pose. These conventional approaches often concentrate on enhancing match quality or eliminating potential outlier correspondences, overlooking the inherent connection between feature matching and pose estimation. Therefore, we propose EIMPNet, an end-to-end optimization network for joint feature matching and pose estimation, generating sparse features and pose simultaneously. Specifically, we use an iterative graph neural network to enhance pre-extracted descriptors, facilitating the network in the extraction of robust matches. We embed differentiable pose estimation into the matching network, along with the match loss for training. This strategy enables the network to iteratively predict matches that are more favourable to pose estimation and naturally reduces the impact of outliers. Experimental results on HPatches and MegaDepth datasets show that our end-to-end training enhances pose estimation metrics and efficiency compared to baseline methods, eliminating the need for RANSAC iterations.