Skip to content

Latest commit

 

History

History
108 lines (75 loc) · 6.1 KB

README.md

File metadata and controls

108 lines (75 loc) · 6.1 KB
Raw

This work is part of the research carried out for the paper here:

Alvarez-Tunon, Olaya, Yury Brodskiy, and Erdal Kayacan. "Loss it right: Euclidean and riemannian metrics in learning-based visual odometry." In ISR Europe 2023; 56th International Symposium on Robotics, pp. 107-111. VDE, 2023.

Quickstart 🦔

Add the base directory of this repo to your pythonpath

export PYTHONPATH=$PYTHONPATH:/your-path/VO_baseline

1. Data loaders

1.1 Generate configs for data loaders (Automatically!)

VO_baseline is designed for an easy deployment in your computer. It has data loaders for KITTI, TUM-RGBD, EuRoC, Aqualoc and MIMIR. This repository expects your to have those datasets under the same folder, and under the names KITTI, TUM, EuRoC, Aqualoc and MIMIR.

If you have it like that, you can go to the file under /scripts/generate_configs.sh This script will automatically generate configs to load the aforementioned datasets. Within that file, you will need to modify the following variables to fit your setup:

# Set conda environment
CONDA_BASE=$(conda info --base)
source "$CONDA_BASE"/etc/profile.d/conda.sh
conda activate olayaenv

After the conda activate directive, put the name of your code environment. Mine is olayaenv.

# Set directories
datasetRoot=$HOME/Datasets
Datasets="KITTI MIMIR Aqualoc/Archaeological_site_sequences EuRoC TUM"

In datasetRoot you need to put the path to your folder containing all datasets. The variable Datasets indicates for which datasets available in the pipeline we want to create configs.

2. Networks 👁️

2.1. DeepVO

Wang, S., Clark, R., Wen, H., & Trigoni, N. (2017, May). Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In 2017 IEEE international conference on robotics and automation (ICRA) (pp. 2043-2050). IEEE.

It is based on a pretrained Flownet. You can find the Flownet weights at : https://github.com/dancelogue/flownet2-pytorch

Under /configs/train you can find several experiment configurations for DeepVO, as depicted in the figure above: deepvo, deepvo_quat and deepvo_se3.

2.1.1 deepvo

Original setup for DeepVO. Here, the output head is a 6-dimensional vector corresponding to the translation vector and the three Euler angles. Under /configs/train/deepvo you can find the configs for the experiments under this setup:

  • original_paper.yml: with data explit and model configuration as proposed in the original DeepVO paper.
  • icra23.yml: with data split as proposed by me in my ICRA workshop paper [link TBD].
  • RWzhou.yml: with data split as proposed in [TBD].

You can train DeepVO for this experiment by executing the script under scripts/train/train_deepvo.py, and choosing the config you want to use.

2.1.2 deepvo_quat

Here, the output head is a 7-dimensional vector corresponding to the translation vector and the quaternion. Under /configs/train/deepvo_quat you can find the configs for the experiments under this setup:

  • original_paper.yml: with data explit and model configuration as proposed in the original DeepVO paper.
  • icra23.yml: with data split as proposed by me in my ICRA workshop paper [link TBD]. The loss function corresponds to the Euclidean loss
  • icra23_geodesic.yml: with data split as proposed in [TBD]. The loss function corresponds to the geodesic loss.

You can train DeepVO for this experiment by executing the script under scripts/train/train_deepvo_quat.py, and choosing the config you want to use.

2.1.3. deepvo_se3

Original setup for DeepVO. Here, the output head is a 6-dimensional vector corresponding to the translation vector and the three Euler angles. Under /configs/train/deepvo you can find the configs for the experiments under this setup:

  • original_paper.yml: with data explit and model configuration as proposed in the original DeepVO paper.
  • icra23.yml: with data split as proposed by me in my ICRA workshop paper [link TBD].
  • RWzhou.yml: with data split as proposed in [TBD].

You can train DeepVO for this experiment by executing the script under scripts/train/train_deepvo_se3.py, and choosing the config you want to use.

All the DeepVO models have been trained under the same experimental setup as follows:

Train Validation Test
00, 02-06,08 07,09 01,10
  • Optimiser: Adgrad
    • Epochs: 200
    • learning rate: 0.001
    • Stacking two images (seq length = 2)

3. Inference

So you have your model and now want to see which results it's inferring? For that, you can execute the script under visualization/<algorithm_name>_inference_results.py. The inference scripts that are available so far are:

  • DeepVO

    visualization/deepvo_inference_results.py. For this script, you need to set the variables:

    • test_sequences: a list of all the sequences that you want to infer on.
    • experiment: name of your experiment. This will be used to save the results.
    • models: list of models you want to infer with. The results (that is, the obtained trajectories) will be saved under visualization/<experiment-name> with the naming convention <sequence_name>_<model-name>.csv. Additionally, the trajectories are also saved in the tum format compatible with the evo tools as evo_<sequence_name>_<model-name>.csv.

Acknowledgements

REMARO Logo

This work is part of the Reliable AI for Marine Robotics (REMARO) Project. For more info, please visit: https://remaro.eu/


EU Flag

This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 956200.