MAC-VO Selected as ICRA 2025 Best Paper Award Finalist

We are thrilled to announce that our paper “MAC-VO: Metrics-aware Covariance for Learning-based Stereo Visual Odometry”, has been selected as a finalist for the Best Perception Paper Award at the upcoming IEEE International Conference on Robotics and Automation (ICRA) 2025.

MAC-VO introduces a novel stereo visual odometry framework that leverages learned uncertainty to enhance keypoint selection and pose graph optimization. Unlike traditional methods that priortize texture-rich features, we proposed a novel metrics-aware covariance model to capture spatial erros and inter-axis correlations, effectively filtering out low-quality features and improving pose estimation accuracy. MAC-VO also enables uncertainty-aware dense mapping without the need for bundle adjustment or multi-frame optimization, facilitating the future research on planning and decision making.

We conducted extensive evaluations on public benchmark datasets, such as VBR, EuRoC, and TartanAir, and our own collection, encompassing a wide range of environments, including indoor, outdoor, and scenarios with extreme lightining conditions. These tests demonstrate that MAC-VO outperforms existing visual odometry algorithms and even some SLAM systems in difficult scenarios.

We look forward to presenting our work at ICRA 2025 in Atlanta, Georgia, on Tuesday, May 20th at 4:30 PM inRoom 302. For more details, please visit our project website and paper.