D 2-Net : A Trainable CNN for Joint Description and Detection of Local Features – Supplementary Material

This supplementary material provides the following additional information: Section 1 details how we chose the threshold for Lowe’s ratio test [5] used for the 3D reconstructions in Section 5.2 in the paper. As mentioned in Section 4.3 in the paper, Section 2 provides implementation details on the architecture. In addition, the section also evaluates another backbone architecture (ResNet [3]). Section 3 provides additional details on the loss function used to train our method. Section 4 shows qualitative examples for the matches found with our approach on the InLoc [14] and Aachen Day-Night [8, 9] datasets.