Joint Out-of-Distribution Detection and Uncertainty Estimation for Trajectory Prediction

This is the official project page including the paper and the code of JOODU:

Joint Out-of-Distribution Detection and Uncertainty Estimation for Trajectory Prediction
Julian Wiederer, Julian Schmidt, Ulrich Kressel, Klaus Dietmayer, Vasileios Belagiannis
Accepted for presentation at IROS 2023 in Detroit, MI, US.

Citation

Please cite our work, if you use our source code:

@InProceedings{wiederer2023joodu,
  author={Julian Wiederer and Julian Schmidt and Ulrich Kressel and Klaus Dietmayer and Vasileios Belagiannis},
  booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Sytems (IROS)}, 
  title={Joint Out-of-Distribution Detection and Uncertainty Estimation for Trajectory Prediction}, 
  year={2023}}

License

JOODU is licensed under Creative Commons Attribution-NonCommercial 4.0 International License.

Check LICENSE for more information.

Installation

The following section describes the setup of the environment and the installation of the required packges including the Shifts and the Argoverse API.

Conda Environment

We recommend using Anaconda. The installation is described on the following page:
https://docs.anaconda.com/anaconda/install/linux/

Next, install the required packages

conda env create -f environment.yml

and activate the environment

conda activate joodu

Install Shifts API

Install the API for the Shifts Vehicle Motion Prediction Dataset as described here:
https://github.com/Shifts-Project/shifts/tree/main/sdc

Install Argoverse API

The argoverse-api is used to convert the HD-map provided by Shifts into the Argoverse format, which is consumed by the trajectory prediction model.

pip install git+https://github.com/argoai/argoverse-api.git

Setup Shifts Dataset

Donwload and extract the datasets

chmod +x scripts/fetch_shifts_dataset.sh
./scripts/fetch_shifts_dataset.sh

The data is pre-processed online during the evaluation.

Model

Download the model parameters from google drive by running

chmod +x scripts/download_model.sh
./download_model.sh

Train the Model

The following command runs the two-stage training on the Shifts train set.
In case you run the training for the first time, the data pre-processing is performed online.

python train.py

In case you would like to skip the first or the second stage, respectively, use the CLI commands below.

Train the first stage only, i.e. only the trajectory prediction model is trained:

python train.py --skip_ood_model_training --skip_uncertainty_model_training

Train the second stage only, i.e. only the OOD detection model and uncertainty estimation model are trained given an existing trajectory prediction model checkpoint under <traj_pred_model_path>:

python train.py --traj_pred_model_path <traj_pred_model_path>

For example you can use the checkpoint from the IROS publication

python train.py --traj_pred_model_path /experiments/traj_pred/000_enc_dec/model/enc_dec_iros_2023.ckpt

Test the Model

The following command runs the evaluation on the Shifts eval set.
In case you run the evaluation for the first time, the data pre-processing is performed online.

chmod +x scripts/eval.sh
./scripts/eval.sh