A unified feature-motion consistency framework for robust image matching

Abstract

Establishing reliable feature matches between a pair of images in various scenarios is a long-standing open problem in photogrammetry. Attention-based detector-free matching with coarse-to-fine architecture has been a typical pipeline to build matches, but the cross-attention module with global receptive field may compromise the structural local consistency by introducing irrelevant regions (outliers). Motion field can maintain structural local consistency under the assumption that matches for adjacent features should be spatially proximate. However, motion field can only estimate local displacements between consecutive images and struggle with long-range displacements estimation in large-scale variation scenarios without spatial correlation priors. Moreover, large-scale variations may also disrupt the geometric consistency with the application of mutual nearest neighbor criterion in patch-level matching, making it difficult to recover accurate matches. In this paper, we propose a unified feature-motion consistency framework for robust image matching (MOMA), to maintain structural consistency at both global and local granularity in scale-discrepancy scenarios. MOMA devises a motion consistency-guided dependency range strategy (MDR) in cross attention, aggregating highly relevant regions within the motion consensus-restricted neighborhood to favor true matchable regions. Meanwhile, a unified framework with hierarchical attention structure is established to couple local motion field with global feature correspondence. The motion field provides local consistency constraints in feature aggregation, while feature correspondence provides spatial context prior to improve motion field estimation. To alleviate geometric inconsistency caused by hard nearest neighbor criterion, we propose an adaptive neighbor search (soft) strategy to address scale discrepancy. Extensive experiments on three datasets demonstrate that our method outperforms solid baselines, with AUC improvements of 4.73/4.02/3.34 in two-view pose estimation task at thresholds of 5°/10°/20° on Megadepth test, and 5.94% increase of accuracy at threshold of 1px in homography task on HPatches datasets. Furthermore, in the downstream tasks such as 3D mapping, the 3D models reconstructed using our method on the self-collected SYSU UAV datasets exhibit significant improvement in structural completeness and detail richness, manifesting its high applicability in wide downstream tasks. The code is publicly available at https://github.com/BunnyanChou/MOMA.