Accuracy Evaluation of Homography-Based Camera Pose and Point Depth Calculation for a Deep Learning Dataset

In recent years, large datasets have been provided for DL (Deep Learning) to train the neural network in calculating camera poses and object depths from images. The datasets were typically obtained using RGB-D (Red, Green, Blue, and Depth) cameras, however, which have limitations in capturing depths in many cases. To overcome the limitations of RGB-D cameras, many DL approaches have been proposed to overcome the problems. However, the accuracy potentials of pose and depth calculated by DL approaches have not been studied well. Thus, this study selected the ScanNet dataset for the experiment and presents a method to evaluate the accuracy of pose and depth that can be obtained using point matching and homography. From the pairs of matching points in two views, their fundamental matrix, essential matrix, and camera matrix were reconstructed and then the camera pose and depth were calculated. The experimental results show that the relative rotation angles were calculated accurately by using homography and the translation angles were calculated less accurately than the rotation angles. In addition, the experiments show that the depths of matching points could be calculated with reasonable accuracy by using the camera matrix constructed by matching.