The Structure-from-Motion (SfM) framework leverages detector-free matching to handle texture-poor scenes. The framework consists of a two-stage pipeline: detector-free matching and coarse SfM followed by iterative refinement. In the detector-free matching phase, feature locations are quantized into a coarse grid to improve consistency, and a coarse SfM model is reconstructed. The iterative refinement pipeline alternates between feature track refinement and geometry refinement to enhance pose and point cloud accuracy. The feature track refinement module is based on a transformer-based multi-view matching network, while the geometry refinement module uses bundle adjustment and track topology adjustment.

The detector-free SfM framework eliminates the need for sparse keypoint detection at the beginning of the pipeline, making it more robust in challenging scenarios such as low-textured regions. The framework is scalable to large-scale scenes and can handle extreme viewpoint and illumination changes. By quantizing matches and building a coarse SfM model, the framework provides initial camera poses and scene structures for later refinement. The iterative refinement pipeline improves accuracy by enhancing feature tracks and refining geometry using multi-view information.

The framework additionally introduces a new texture-poor SfM dataset and demonstrates the framework's effectiveness on challenging scenes. The proposed framework outperforms state-of-the-art detector-based SfM systems in terms of various metrics. By leveraging detector-free matching and an iterative refinement pipeline, the framework can recover accurate camera poses with high registration rates even in texture-poor scenes. The transformer-based multi-view matching network enhances the discriminativeness of features by encoding positional and multi-view context, leading to improved accuracy in pose and point cloud reconstruction.