HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention

arXiv | paper | poster\ This repository is the official implementation of HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention published in CVPR 2024. AnoverviewofHPNet

Setup
Datasets
Training
Validation
Testing
Pre-trained Models & Results
Acknowledgements
Citation

Setup

Clone the repository and set up the environment:

git clone https://github.com/XiaolongTang23/HPNet.git
cd HPNet
conda create -n HPNet python=3.8
conda activate HPNet
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install torch_geometric==2.3.1
conda install pytorch-lightning==2.0.3

Note: For compatibility, you may experiment with different versions, e.g., PyTorch 1.12.1 has been confirmed to work.

Datasets

Argoverse

1. Download the [Argoverse Motion Forecasting Dataset v1.1](https://www.argoverse.org/av1.html#download-link). After downloading and extracting the tar.gz files, organize the dataset directory as follows: ``` /path/to/Argoverse_root/ ├── train/ │ └── data/ │ ├── 1.csv │ ├── 2.csv │ ├── ... └── val/ └── data/ ├── 1.csv ├── 2.csv ├── ... ``` 2. Install the [Argoverse API](https://github.com/argoverse/argoverse-api).

INTERACTION

1. Download the [INTERACTION Dataset v1.2](https://interaction-dataset.com/). Here, we only need the data for the multi-agent tracks. After downloading and extracting the zip files, organize the dataset directory as follows: ``` /path/to/INTERACTION_root/ ├── maps/ ├── test_conditional-multi-agent/ ├── test_multi-agent/ ├── train/ │ ├── DR_CHN_Merging_ZS0_train │ ├── ... └── val/ ├── DR_CHN_Merging_ZS0_val ├── ... ``` 2. Install the map dependency [lanelet2](https://github.com/fzi-forschungszentrum-informatik/Lanelet2): ``` pip install lanelet2==1.2.1 ```

Training

Data preprocessing may take several hours the first time you run this project. Training on 8 RTX 4090 GPUs, one epoch takes about 30 and 6 minutes for Argoverse and INTERACTION, respectively.

# For Argoverse
python HPNet-Argoverse/train.py --root /path/to/Argoverse_root/ --train_batch_size 2 --val_batch_size 2 --devices 8

# For INTERACTION
python HPNet-INTERACTION/train.py --root /path/to/INTERACTION_root/ --train_batch_size 2 --val_batch_size 2 --devices 8

Validation

# For Argoverse
python HPNet-Argoverse/val.py --root /path/to/Argoverse_root/ --val_batch_size 2 --devices 8 --ckpt_path /path/to/checkpoint.ckpt

# For INTERACTION
python HPNet-INTERACTION/val.py --root /path/to/INTERACTION_root/ --val_batch_size 2 --devices 8 --ckpt_path /path/to/checkpoint.ckpt

Testing

# For Argoverse
python HPNet-Argoverse/test.py --root /path/to/Argoverse_root/ --test_batch_size 2 --devices 1 --ckpt_path /path/to/checkpoint.ckpt

# For INTERACTION
python HPNet-INTERACTION/test.py --root /path/to/INTERACTION_root/ --test_batch_size 2 --devices 1 --ckpt_path /path/to/checkpoint.ckpt

Pre-trained Models & Results

Argoverse

Pre-trained model: Download here
Performance Metrics:

Split	brier-minFDE	minFDE	MR	minADE
Val	1.5060	0.8708	0.0685	0.6378
Test	1.7375	1.0986	0.1067	0.7612

INTERACTION

Pre-trained model: Download here
Performance Metrics:

Split	minJointFDE	minJointADE
Val	0.5577	0.1739
Test	0.8231	0.2548

Acknowledgements

We sincerely appreciate Argoverse, INTERACTION,QCNet and HiVT for their awesome codebases.

Citation

If HPNet has been helpful in your research, please consider citing our work:

@inproceedings{tang2024hpnet,
  title={Hpnet: Dynamic trajectory forecasting with historical prediction attention},
  author={Tang, Xiaolong and Kan, Meina and Shan, Shiguang and Ji, Zhilong and Bai, Jinfeng and Chen, Xilin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={15261--15270},
  year={2024}
}

XiaolongTang23 / HPNet

readme