Comparison of Computer Vision and Photogrammetric Approaches for Epipolar Resampling of Image Sequence

Abstract

Epipolar resampling is the procedure of eliminating vertical disparity between stereo images. Due to its importance, many methods have been developed in the computer vision and photogrammetry field. However, we argue that epipolar resampling of image sequences, instead of a single pair, has not been studied thoroughly. In this paper, we compare epipolar resampling methods developed in both fields for handling image sequences. Firstly we briefly review the uncalibrated and calibrated epipolar resampling methods developed in computer vision and photogrammetric epipolar resampling methods. While it is well known that epipolar resampling methods developed in computer vision and in photogrammetry are mathematically identical, we also point out differences in parameter estimation between them. Secondly, we tested representative resampling methods in both fields and performed an analysis. We showed that for epipolar resampling of a single image pair all uncalibrated and photogrammetric methods tested could be used. More importantly, we also showed that, for image sequences, all methods tested, except the photogrammetric Bayesian method, showed significant variations in epipolar resampling performance. Our results indicate that the Bayesian method is favorable for epipolar resampling of image sequences.

Analysis of Bundle Adjustments and Epipolar Model Accuracy According to Flight Path Characteristics of Uav

“…To improve the accuracy of stereo plotting, the epipolar model accuracy is important. Epipolar modeling is the procedure of eliminating vertical disparity between stereo images (Kim and Kim, 2016). As an ideal epipolar model has yparallax of 0, we can determine epipolar model accuracy with yparallax.…”

Development of a Single-View Odometer Based on Photogrammetric Bundle Adjustment

“…The relative geometry of two images is estimated based on coplanarity conditions in photogrammetry (Kim and Kim, 2016). This is called relative orientation.…”

Extracting three-dimensional (3D) spatial information from sequential oblique unmanned aerial system (UAS) imagery for digital surface modeling

“…Dissimilarity computations on epipolar lines (i.e. horizontal scan lines) are one of the practical approaches (Kim and I.and Kim 2016) to achieve efficiency. In addition, despite the widespread use of start-of-the-art deep learning approaches, such as deep MVS (Huang et al 2018), which can yield better results than SGM, its high dependence on training datasets introduces a large gap with the classical algorithms in terms of universality (Liu et al 2020).…”