Global-Aware Edge Prioritization for Pose Graph Construction in SfM

Tong Wei, Giorgos Tolias, Jiri Matas, Daniel Barath

Keywords: pose graph construction; structure from motion

Abstract: The pose graph is an essential component of Structure-from-Motion (SfM) pipelines, where images form the nodes and edges encode relative poses between them. These graphs are typically sparse to reduce the cost of geometric verification required for each candidate edge. In this paper, we focus on robust pose graph initialization, performed at the very beginning of the SfM pipeline. Traditionally, this step relies on image retrieval methods applied independently to each image, connecting it to the $k$ most similar ones according to, e.g., embedding similarity. While effective in practice, this greedy approach does not allow communication across image pairs during graph construction. We address this limitation through the novel concept of edge prioritization, which ranks edges by their utility for SfM. We achieve this through the following two contributions. First, we propose an image representation network combined with a graph neural network (GNN), trained with SfM-derived supervision to predict edge ranks. The GNN exploits global context from the entire image set to guide pair selection. Second, we introduce an edge selection strategy based on minimum spanning trees, which uses predicted ranks to identify the most promising pairs. By integrating global information at both stages, our approach substantially improves SfM reconstruction in the high-speed regime, particularly when operating with very sparse pose graphs. Code will be released.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2031