Graph-Based Consistent Matching for Structure-from-Motion - Supplementary Material

The initialization of the graph-based matching method depends on a minimum spanning tree (MST) parameterized by edge weights. Although we use the quadratic mean of mutual ranks as the edge weight, Structure-from-Motion (SfM) is not much affected by different weighting strategies, as we will show in the following experiment. It is shown that this method is not sensitive to the initial MST step. We use five different weighting strategies to generate the initial MST and apply our method on the highly ambiguous TempleOfHeaven dataset. Namely, we combine the rank of the i-th image in jth image’s candidate list Rank i (j), and its mutual counterpart Rank 2 j (i) using quadratic mean, harmonic mean, arithmetic mean, geometric mean and maximum respectively. All the five initialization strategies yield a successful reconstruction, without noticeable differences in re-projection error and SfM results (see Table 1 and Figure 1).