Open David-Willo opened 10 months ago
Hi, thanks for the question.
In our experiments, we found that the front-end, when refined on half of the trajectories in the KITTI dataset, demonstrated improved performance on the remaining unseen trajectories. We think this suggests that our learning framework generalizes the VO network to new environments.
Please see Section IV.D of our paper for more details. https://arxiv.org/pdf/2306.07894
Thanks for your interesting work. Is the refined front-end able to do pose estimation in new scenes that are not used in the training process?