This code is a supplementary material to the paper "TLIO: Tight Learned Inertial Odometry". To use the code here requires the user to generate its own dataset and retrain. For more information about the paper and the video materials, please refer to our website.

All dependencies can be installed using conda via

conda env create -f environment.yaml

Then the virtual environment is accessible with:

conda activate tlio

Next commands should be run from this environment.

A dataset is needed in numpy format to run with this code. We have released the dataset used in the paper. The data can be downloaded here or with the following command (with the conda env activated) at the root of the repo:

gdown 14YKW7PsozjHo_EdxivKvumsQB7JMw1eg
mkdir -p local_data/ # or ln -s /path/to/data_drive/ local_data/
unzip golden-new-format-cc-by-nc-with-imus-v1.5.zip -d local_data/
rm golden-new-format-cc-by-nc-with-imus-v1.5.zip

https://drive.google.com/file/d/14YKW7PsozjHo_EdxivKvumsQB7JMw1eg/view?usp=share_link The dataset tree structure looks like this. Assume for the examples we have extracted the data under root directory local_data/tlio_golden:

local_data/tlio_golden
├── 1008221029329889
│   ├── calibration.json
│   ├── imu0_resampled_description.json
│   ├── imu0_resampled.npy
│   └── imu_samples_0.csv
├── 1014753008676428
│   ├── calibration.json
│   ├── imu0_resampled_description.json
│   ├── imu0_resampled.npy
│   └── imu_samples_0.csv
...
├── test_list.txt
├── train_list.txt
└── val_list.txt

imu0_resampled.npy contains calibrated IMU data and processed VIO ground truth data. imu0_resampled_description.json describes what the different columns in the data are. The test sequences contain imu_samples_0.csv which is the raw IMU data for running the filter. calibration.json contains the offline calibration. Attitude filter data is not included with the release.

There are three different modes for the network part.--mode parameter defines the behaviour of main_net.py. Select between train, test
train: training a network model with training and validation dataset.
test: running an existing network model on testing dataset to obtain concatenated trajectories and metrics. \

Parameters:

--root_dir: dataset root directory. Each subfolder of root directory is a dataset.
--out_dir: training output directory, where checkpoints and logs folders will be created to store trained models and tensorboard logs respectively. A parameters.json file will also be saved.

Example:

python3 src/main_net.py \
--mode train \
--root_dir local_data/tlio_golden \
--out_dir models/resnet \
--epochs 100

Note: We offer multiple types of dataloaders to help speed up training. The --dataset_style arg can be ram, mmap, or iter. ram stores all the sequences in RAM, mmap uses memmapping to only keep part of the sequences in RAM at once, and iter is an iterable-style dataloader for larger datasets, which sacrifices true randomness. The default is mmap, which offers the best tradeoff, and typically works for the dataset provided. However, we found that on server-style machines, the memmapping can cause RAM to fill up for some reason (it seems to work best on personal desktops). If the training is getting killed by your OS or taking up too much RAM, you may try setting --workers to 1, --dataset_style to ram, and/or --no-persistent_workers.

tensorboard --logdir models/resnet/logs/

Parameters:

--model_path: path of the trained model to test with.
--out_dir: testing output directory, where a folder for each dataset tested will be created containing estimated trajectory as trajectory.txt and plots if specified. metrics.json contains the statistics for each dataset.

Example:

python3 src/main_net.py \
--mode test \
--root_dir local_data/tlio_golden \
--model_path models/resnet/checkpoint_*.pt \
--out_dir test_outputs

Warning: network testing use the ground truth orientations for displacement integration. Please do not consider them as benchmarks, they are more like a debugging tool.

Batch scripts are under src/batch_analysis module. Execute batch scripts from the src folder.

Batch testing tests a list of datasets using multiple models and for each model save the trajectories, plots and metrics into a separate model folder. Output tree structure looks like this:

batch_test_outputs
├── model1
│   ├── seq1
│   │   ├── trajectory.txt
│   │   └── *.png
│   ├── seq2
...
│   └── metrics.json
├── model2
│   ├── seq1
...
│   └── metrics.json
...

Create an output directory and go to the src folder

mkdir batch_test_outputs
cd src

Run batch tests. --model_globbing is the globbing pattern to find all models to test. Here we only have one.

python -m batch_runner.net_test_batch \
--root_dir ../local_data/tlio_golden \
--model_globbing "../models/*/checkpoint_best.pt" \
--out_dir ../batch_test_outputs \
--save_plot

If you saved plot, you can visualize there:

feh ../	batch_test_outputs/models-resnet/*/view.png # example using the `feh` image visualizer, use your favorite one

After running testing and evaluation in batches, the statistics are saved in either metrics.json. To visualize the results and compare between models, we provide scripts that display the results in an interactive shell through iPython. The scripts are under src/analysis module.

To visualize network testing results from metrics.json including trajectory metrics and testing losses, go to src folder and run

python -m analysis.display_json \
--glob_dataset "../batch_test_outputs/*/"

This will leave you in an interactive shell with a preloaded panda DataFrame d. You can use it to visualize all metrics with the following helper function:

plot_all_stats_net(d)

The EKF expects the model to be in torchscript format.

Example: From the repo root:

python3 src/convert_model_to_torchscript.py \
--model_path models/resnet/checkpoint_best.pt \
--model_param_path models/resnet/parameters.json \
--out_dir models/resnet/

which will create models/resnet/model_torchscript.pt.

Use src/main_filter.py for running the filter and parsing parameters. The program supports running multiple datasets on one specified network model.

Parameters:

--model_path: path to saved model checkpoint file.
--model_param_path: path to parameter json file for this model.
--out_dir: filter output directory. This will include a parameters.json file with filter parameters, and a folder for each dataset containing the logged states, default to not_vio_state.txt.
--erase_old_log: overwrite old log files. If set to --no-erase_old_log, the program would skip running on the datasets if the output file already exists in the output directory.
--save_as_npy: convert the output txt file to npy file and append file extension (e.g. not_vio_state.txt.npy) to save space.
--initialize_with_offline_calib: initialize with offline calibration of the IMU. If set to --no-initialize_with_offline_calib the initial IMU biases will be initialized to 0.
--visualize: if set, open up an Open3D window to visualize the filter running. This is of course optional.

Example:

python3 src/main_filter.py \
--root_dir local_data/tlio_golden \
--model_path models/resnet/model_torchscript.pt \
--model_param_path models/resnet/parameters.json \
--out_dir filter_outputs \
--erase_old_log \
--save_as_npy \
--initialize_with_offline_calib \
--dataset_number 22 \
--visualize

Please refer to main_filter.py for a full list of parameters.

Batch script batch_runner/filter_batch provides functionality to run the main file in batch settings. Go to src folder to run the module and you can set the parameters to test within the script (e.g. different update frequencies).

Example:

cd src
python -m batch_runner.filter_batch \
--root_dir ../local_data/tlio_golden \
--model_globbing "../models/*/model_torchscript.pt" \
--out_dir ../batch_filter_outputs

To generate plots of the states of the filter and to generate metrics.json file for both the filter and network concatenation approaches, batch run plot_state.py on the existing filter and network testing outputs.

Parameters:

--runname_globbing: globbing pattern for all the model names to plot. This pattern should match between filter and ronin and exist in both --filter_dir and --ronin_dir.
--no_make_plots: not to save plots. If removed plots will be saved in the filter output folders for each trajectory.

Example:

python -m batch_runner.plot_batch \
--root_dir ../local_data/tlio_golden \
--runname_globbing "*" \
--filter_dir ../batch_filter_outputs_uf20 \
--ronin_dir ../batch_test_outputs

Up to now a metrics.json file will be added to each model folder, and the tree structure would look like this:

batch_filter_outputs
├── model1
│   ├── seq1
│   │   ├── *.png
│   │   ├── not_vio_state.txt.npy
│   │   └── vio_states.npy
│   ├── seq1
│   │   ├── *.png
│   │   ├── not_vio_state.txt.npy
│   │   └── vio_states.npy
...
│   ├── metrics.json
│   └── parameters.json
├── model2
...

Visualize the plot from the filter and ronin:

feh ../batch_filter_outputs_uf20/models-resnet/*/position-2d.png # example using the `feh` image visualizer, use your favorite one

To generate plots from the metrics:

python -m analysis.display_json \
--glob_dataset "../batch_filter_outputs_uf20/*/"

# then run in the interactive session
# plot_sysperf_cdf(d)