Learning to Detect Scene Landmarks for Camera Localization

In this supplementary document, we present additional quantitative results that could not be included in the main paper. We show extensive qualitative results from our method on the INDOOR-6 dataset and also include a supplemental video. Finally, we discuss some failure cases. Quantitative Results In this section, we show the storage efficiency of our method (NBE+SLD) compared to a retrieval and matchingbased method (HLoc [3]) (Section 1.1). Next, we further compare accuracy between our method and multiple baselines through a recall plot that uses a range of thresholds (Section 1.2). Finally, we report bearing errors for predicted landmarks on INDOOR-6 and 7-SCENES [6] datasets (Section 1.3). Storage comparison NBE+SLD requires constant storage. Figure 1 reports the storage requirements for HLoc and our method for each scene in the INDOOR-6 dataset. Our method requires 0.135 GB of storage for the SLD and NBE networks' parameters that are constant for all the scenes. This is significantly smaller than HLoc that requires 1.5GB and 1.2GB on the two larger scenes -scene1 and scene5, respectively. HLoc stores SuperPoint [1] and SuperGlue [4] networks' parameters and SuperPoint's features and VLAD [2] image descriptors for all the database images. The storage for features grows linearly with the number of database images, and that can dominate total storage on large scenes.