Learning generalized visual odometry using position-aware optical flow and geometric bundle adjustment

Recent visual odometry (VO) methods incorporating geometric algorithm into deep-learning architecture have shown outstanding performance on the challenging monocular VO task. Despite encouraging results are shown, previous ignore requirement of generalization capability under noisy environment and various scenes. To address this issue, work first proposes a novel optical flow network (PANet). Compared with that predict as direct regression task, our PANet computes by predicting it discrete position space probability volume, then converting to flow. Next, we improve bundle adjustment module fit self-supervised training pipeline introducing multiple sampling, ego-motion initialization, dynamic damping factor adjustment, Jacobi matrix weighting. In addition, normalized photometric loss function is advanced depth estimation accuracy. The experiments show proposed system not only achieves comparable other state-of-the-art learning-based KITTI dataset, but also significantly improves compared geometry-based, hybrid systems outdoor (KAIST)