This repository corresponds to the official source code of the ICCV 2021 paper:

On Exposing the Challenging Long Tail in Future Prediction of Traffic Actors

We use the same requirements as the Trajectron++, see: https://github.com/StanfordASL/Trajectron-plus-plus

Additionally, it is essential to download the Trajectron++ code, rename it to Trajectron_plus_plus and place it next to other folders (e.g., data/, models/).

The test data files are provided under data/. These are the result of running the processing script of the Trajectron++, see: https://github.com/StanfordASL/Trajectron-plus-plus/blob/master/experiments/pedestrians/process_data.py

For the processed files, you can run the processing script of nuScenes at: https://github.com/StanfordASL/Trajectron-plus-plus/blob/master/experiments/nuScenes/process_data.py

All pretrained models (EWTA and with contrastive learning) are provided under models/.

This is an example call of the testing script (test Trajectron++EWTA on ETH):

python test.py --model models/eth_ewta/ --checkpoint 490 --data data/eth_test.pkl --kalman kalman/eth_PEDESTRIAN_test_kalman.pkl --node_type PEDESTRIAN

Another example to test all vehicles on nuScenes dataset:

python test.py --model models/nuScenes_ewta/ --checkpoint 25 --data data/nuScenes_test_full.pkl --kalman kalman/nuScenes_VEHICLE_test_kalman.pkl --node_type VEHICLE

Coming soon...

If you use our repository or find it useful in your research, please cite the following paper:

@InProceedings{MCMB21,
  author       = "O. Makansi and {\"O}. {\c{C}}i{\c{c}}ek and Y. Marrakchi and T. Brox",
  title        = "On Exposing the Challenging Long Tail in Future Prediction of Traffic Actors",
  booktitle    = "IEEE International Conference on Computer Vision (ICCV)",
  month        = " ",
  year         = "2021",
  url          = "http://lmb.informatik.uni-freiburg.de/Publications/2021/MCMB21"
}