Answers are grounded in this paper and cite their sources.
Demonstrates improved accuracy in pose estimation using learning-based outlier rejection in visual odometry.
Visual odometry (VO) estimates the relative motion of a camera from sequential images and is widely used in autonomous driving and mobile robotics. Conventional feature-based VO methods typically employ Random Sample Consensus (RANSAC) for outlier rejection; however, its reliance on iterative random sampling makes it sensitive to large scene changes and high outlier ratios. To address this limitation, we propose a feature-based stereo VO framework that replaces the conventional outlier rejection module with NCMNet, a learning-based alternative. By leveraging learned correspondence filtering, the proposed method improves robustness without relying on random sampling. We evaluate the proposed approach on sequences 09 and 10 of the KITTI Odometry dataset by comparing it with a RANSAC-based baseline under identical settings. The proposed method consistently reduces both Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). In particular, the ATE on sequence 09 decreases from 36.538 m to 17.713 m. These results demonstrate that learning-based outlier rejection provides a more robust alternative to conventional methods, leading to improved pose estimation accuracy in feature-based VO.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: