FEND

Code for CVPR 2023 'FEND: A Future Enhanced Distribution-Aware Contrastive Learning Framework for Long-tail Trajectory Prediction'

Requirements

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

Environment Setup

Create a conda environment.

conda create --name trajectron++ python=3.9 -y
source activate trajectron++
pip install -r requirements.txt

Data

nuScenes (Bird's-eye view):

For the processed files, you can run the processing script of nuScenes at:

python Trajectron_plus_plus/experiments/nuScenes/process_data.py

Pre-trained Models

Pretrained models are provided under models/.

Testing

Example to test our model on nuScenes dataset:

python test_nuscenes_fend.py --model models/nuScenes_model/trajectron_map_int_fend_ewta_withoutfrequency --checkpoint 25 --data data/nuScenes_test_full.pkl --kalman kalman/nuScenes_VEHICLE_test_ewta_baseline_withoutfrequency.pkl --node_type VEHICLE

Example to test the baseline traj++ewta model on nuScenes dataset:

python test_nuscenes_fend.py --model models/nuScenes_model/trajectron_map_int_fend_ewta_withoutfrequency --checkpoint 25 --data data/nuScenes_test_full.pkl --kalman kalman/nuScenes_VEHICLE_test_ewta_baseline_withoutfrequency.pkl --node_type VEHICLE

We use the Traj++EWTA without resampling as the baseline model to select the hard samples, as our method is implemented on it. Testing with the Traj++EWTA with resampling as the baseline model to select hard samples can be done as follows:

python test_nuscenes_fend.py --model models/nuScenes_model/trajectron_map_int_fend_ewta_withoutfrequency --checkpoint 25 --data data/nuScenes_test_full.pkl --kalman kalman/nuScenes_VEHICLE_test_ewta_baseline_withfrequency.pkl --node_type VEHICLE

Training

Example to train our model on nuScenes dataset:

python Trajectron_plus_plus/trajectron/train_nuscenes_fend.py

Citation

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

@inproceedings{wang2023fend,
  title={Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction},
  author={Wang, Yuning and Zhang, Pu and Bai, Lei and Xue, Jianru},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={1400--1409},
  year={2023}
}