GaLAM: Two-Stage Outlier Detection Algorithm

Image matching aims to establish correspondences between images and estimate relative transformations, serving as a critical component in many geometric computer vision tasks. However, due to the limitations of descriptors and the presence of moving objects in the target region, results from nearest-neighbor matching often include numerous outliers. Recently, significant efforts have focused on effectively filtering these outliers to improve image matching performance. In this paper, we propose Globally-and-Locally-Affine Matching (GaLAM), a two-stage outlier detection method. In the first stage, GaLAM optimizes the edge features of Adaptive Locally Affine Matching (AdaLAM) and replaces the statistical neighborhood-based approach with classical geometric deviation metrics. In the second stage, Progressive Sample Consensus (PROSAC) is used to sample seed points that align with the fundamental matrix, followed by filtering outliers based on projection deviation. This approach enhances the model’s robustness in complex scenes, particularly in scenarios with moving objects in the target region. We evaluated GaLAM on diverse datasets, and the results demonstrate that our method achieves significant improvements over AdaLAM in most scenarios.