Diverse and Admissible Trajectory Forecasting through Multimodal Context Understanding
This code is PyTorch implementation of our work, Diverse and Admissible Trajectory Forecasting through Multimodal Context Understanding (Seong Hyeon Park, Gyubok Lee, Manoj Bhat, Jimin Seo, Minseok Kang, Jonathan Francis, Ashwin R. Jadhav, Paul Pu Liang and Louis-Philippe Morency).
Trajectory forecasting task was implemented using normalizing-flow, and achieving diverse while admissible trajectorys for vehicles.
Dataset
You will have to download dataset into data/[corresponding dataset], then verify the subdirectory of each dataset. Dataset should have structure same as:
-[dataset]
|-
Dataset will be extracted as cache at the initial execution. When not specified, cache file will be used for preceeding experiments.
Training
See details on train.py
To train proposed method;
python setup.py develop && CUDA_VISIBLE_DEVICES=2 python train.py --config=config_atgs_cam_nf
Testing
Testing will be used by assigning checkpoint to argument --test_ckpt; Make sure to change the TRAIN flag on the config file to false
python setup.py develop && CUDA_VISIBLE_DEVICES=2 python test.py --config=config_atgs_cam_nf --test_ckpt="xyz.tar"
Things to do
-
Select appropriate License; currently we used GPLv3. - MATF_GAN had runtime error which has fixed. For coherence, this will be updated after recieving it.
- Check the requirements
Citation
Please cite the original publication;
@article{park2020diverse,
title={Diverse and Admissible Trajectory Forecasting through Multimodal Context Understanding},
author={Park, Seong Hyeon and Lee, Gyubok and Bhat, Manoj and Seo, Jimin and Kang, Minseok and Francis, Jonathan and Jadhav, Ashwin R and Liang, Paul Pu and Morency, Louis-Philippe},
journal={arXiv preprint arXiv:2003.03212},
year={2020}
}
License
This code is published under the General Public License version 3.0.