Evaluating the synergy of hand-crafted and AI-driven feature matching in Structure-from-Motion 3D Reconstruction

Abstract. Matching images containing repetitive patterns or uniform textures remains a persistent challenge in structure-from-motion (SfM) 3D reconstruction, as the quality of feature correspondences directly influences camera pose estimation and 3D point cloud generation. Robust feature matching is therefore essential for achieving reliable SfM results. This study investigates the impact of integrating traditional hand-crafted and AI-driven feature-matching approaches on SfM reconstruction performance. SIFT and SuperPoint are selected as representative feature extractors and combined with multiple matching algorithms, including FLANN, LightGlue, and StereoGlue, to evaluate their effectiveness. The resulting feature matches are assessed using three quantitative indicators: (1) the number of successfully matched image pairs, (2) the number of correctly aligned images, and (3) positional errors of the SfM reconstruction relative to GNSS meta-information. The experimental results demonstrate that the combination of SIFT and LightGlue consistently yields the highest number of correctly aligned images and the lowest positional errors across all test cases. StereoGlue produces partially reliable matches; however, its performance is sensitive to the characteristics of the training data. In contrast, while the SIFT–FLANN combination generates a large number of matched pairs, it frequently results in misalignments due to false correspondences in scenes with uniform textures. These findings indicate that coupling hand-crafted feature extraction with AI-assisted matching strategies provides a robust and effective solution for improving SfM reconstruction in challenging imagery datasets.