Figure 1
From: Relative pose estimation from panoramic images using a hybrid neural network architecture
Comparison of MinWarping with neural network image preprocessing (proposed method, right) against MinWarping with edge filtering (left). Example image pairs show a home position (“snapshot”) and a changed position after robot movement (“current view”). All images are represented in RGB color space. Homing examples show a \(7\times 7\) section of a grid image dataset with the snapshot position marked with a red dot and reference homing vectors as red arrows. Gray arrows are overlayed to show the homing results for each current view with respect to the snapshot, considering all illumination variants present in the dataset and positions from ground truth data. All data is from setting “Computer Lab”; see “Dataset”.
