Figure 17
From: Relative pose estimation from panoramic images using a hybrid neural network architecture
Searching columns with minimal distance in the scale plane stack in a situation with occlusion and image regions with few details (taken from the Long Hallway setting). For each subplot, we display a single scale plane with unscaled snapshot and current view (\(\sigma = 1.0\)). We remove the distortion of the scale plane defined by the index shift described in the section “Extensions to the MinWarping algorithm” such that each distance entry aligns with the columns of the images displayed on the edges. The images displayed on the side correspond to the snapshot (top) and current view (left). Both the original as well as the preprocessed images are shown. We mark visible regions with green boxes in the images and with a green tint in the scale plane. Red markers of images and scale plane show columns that are not visible and therefore cannot be matched between snapshot and current view. We display the position and importance of minima with an overlayed scatterplot. The importance of a minimum is calculated by normalizing all minima that contribute to the movement parameters estimated by MinWarping. The highest minima are displayed with small, yellow points; the lowest minima are represented with large, red points. Values in between interpolate simultaneously in size and color. The angular error for MinWarping NP is \( \sim \)7.73°, and for MinWarping EF \( \sim \)93.85°.
