E-VIO

This repository is the official implementation of the papers E-VIO: A Continual Evolving Visual-Inertial Odometry for Drones in Flight, which has been submitted to ISPRS Journal of Photogrammetry and Remote Sensing.

Setup

Installation

• Create conda environment
• Install g2opy, torch==1.10.0, torchvision==0.11.1

Data preparation

To re-train or run the experiments from our paper, please download and pre-process the respective datasets.

Cityscapes

Download the following files

• leftImg8bit_sequence_trainvaltest.zip
• timestamp_sequence.zip
• vehicle_sequence.zip

Oxford RobotCar

Download the following files

• 22015-10-29-12-18-17_stereo_centre.tar, 2015-10-29-12-18-17_gps.tar
• 2015-02-03-08-45-10_stereo_centre.tar, 2015-02-03-08-45-10_gps.tar
• 2015-08-21-10-40-24_stereo_centre.tar, 2015-08-21-10-40-24_gps.tar

Undistort the center images:

python datasets/robotcar.py <IMG_PATH> <MODELS_PATH>

KITTI

Download the KITTI Odometry dataset

• odometry data set
• odometry ground truth poses

Extract the raw data matching the odometry dataset. | 04 | 2011_09_30_drive_0016 | | 09 | 2011_09_30_drive_0033 | | 10 | 2011_09_30_drive_0034 |

EuRoc

Download the EuRoc MAV dataset

• MH_03, MH_05, V2_02

Running the Code

Pre-training

We pre-trained CoVIO on the Cityscapes Dataset.

python main_pretrain.py

Continual Learning with E-VIO

For continual learning, we used the KITTI Odometry Dataset, the Oxford RobotCar Dataset and the EuRoc MAC Dataset. Then run:

python main_adapt.py

License

For academic usage, the code is released under the GPLv3 license. For any commercial purpose, please contact the authors.

Reference

[1] Continual SLAM: Beyond Lifelong Simultaneous Localization and Mapping Through Continual Learning [2] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation