Filtering 2D-3D Outliers by Camera Adjustment for Visual Odometry

Abstract

We study the problem of the discrepancy between model predictions and image measurements in the form of keypoint locations for perspective cameras. In this process, the prediction is made by projecting given 3D points using the known pose of a calibrated camera. We test whether some small camera pose adjustment exists for each measurement such that the mentioned discrepancy vanishes. Such adjustment would allow us to quantify the effect of each measurement on the camera pose. In this paper, we show for the first time that the pose influence assessment of individual measurements can be used to select a subset of the correspondences for accurate 3D triangulation from two views. We further demonstrate via several experiments that the obtained 3D points are well suited to the task of absolute localization. When the 3D points are provided from an anonymized source, the proposed method also selects a suitable subset of 3D points for accurate localization around an initial guess. The long-term effectiveness of our filtration method is demonstrated by integrating the method within a typical framework of visual odometry. The proposed method is evaluated on ETH3D and EuRoC benchmarks with real-world data. The results indicate that the proposed method outperforms the state-of-the-art methods in terms of the point uncertainty measure and camera pose estimation accuracy.