This is the official code for the paper ExAgt: Expert-guided Augmentation for Representation Learning of Traffic Scenarios published in IEEE ITSC 2022, Macau, China
Dataset
- The Argoverse motion prediction dataset is pre-processed and the augmentation and occcupancy grids information are generated. Please download the dataset from the link https://faubox.rrze.uni-erlangen.de/getlink/fiKwbBsaqtsNGPGU1eHmiF/Argoverse.zip>
- Place the dataset in a folder named data
Code
- The file
main.pyis the config file to start the experiment - The main config in the file is the loss_type to be used and the eval metric. The user can change these two values depending on the task under consideration https://github.com/lab176344/ExAgt_Work/blob/d758e4b45fb82b7a4a6ae96fec0a0655a9cd97c2/main.py#L82 and https://github.com/lab176344/ExAgt_Work/blob/d758e4b45fb82b7a4a6ae96fec0a0655a9cd97c2/main.py#L83
- The main file will also start the tensorboard visualisation of the training
- The setting parameters for the training and augmentation are explained in the paper, the user can reproduce the settings
Please cite our paper if you find the code useful
@article{balasubramanian2022exagt,
title={ExAgt: Expert-guided Augmentation for Representation Learning of Traffic Scenarios},
author={Balasubramanian, Lakshman and Wurst, Jonas and Egolf, Robin and Botsch, Michael and Utschick, Wolfgang and Deng, Ke},
journal={arXiv preprint arXiv:2207.08609},
year={2022}
}
If you have any issues with the code or any doubts raise an issue in the repo, we will try to resolve it as soon as possible