ClementPinard
commented
5 years ago Reported results are trained on Eigen's split. I only now realize that odemetry results should be reported with a training on odom split ! Thanks for that ! I'll come up with a patch and maybe a pretrained network to solve this.
In the mean time, you can use author's code, with stacked_sequence_folders instead of sequence_folders for dataset.
For KITTI, it seems that the data preparation code only works for Eigen's split (for depth estimation), but not for Odom split (for pose estimation). I wonder if you can also share your data preparation code for Odom split. Also, I wonder if the reported results on pose are trained on Eigen's split. Thanks.