TrajNet++ Model Zoo

TrajNet++ is an open source framework for training trajectory forecasting models.

The TrajNet++ Model Zoo is a collection of pre-trained, state-of-the-art models for the TrajNet++ challenge. Accompanying each model is the description of model training and the command to train the model.

Trajnetplusplusbaselines Models

This collection of models can be directly trained using the TrajNet++ repository.

Interaction Encoders

Interaction Class	Reference	Command	Description
Occupancy LSTM	Alahi et al.	sh scripts/interaction/occupancy.sh	Occupancy Grid
Social LSTM	Alahi et al.	sh scripts/interaction/social.sh	Social Grid
Directional LSTM	Kothari et al.	sh scripts/interaction/directional.sh	Directional Grid

External Models

This collection of models can be trained using modifications of other open-source model repositories to TrajNet++ format.

Model	Reference	Command	Repository
Trajectory Transformer	Giuliari et al.	./trajnet_scripts/trajnet_run.sh	repository link
STGAT	Huang et al.	./trajnet_scripts/trajnet_run.sh	repository link
Social-STGCNN	Mohamed et al.	./trajnet_scripts/trajnet_run.sh	repository link
CausalHTP	Chen et al.	./trajnet_scripts/trajnet_run.sh	repository link
SocialWays	Amirian et al.	./trajnet_scripts/trajnet_run.sh	repository link
SR-LSTM	Zhang et al.	./trajnet_scripts/trajnet_run.sh	repository link
DAG-Net	Monti et al.	./trajnet_scripts/trajnet_run.sh	repository link

Results

The folowing numbers are obtained using the officially released hyperparameters in the respective repositories. We are still in the process of tuning hyperparameters to optimize performance on TrajNet++

Model	ADE/FDE	Top-3 ADE/FDE	Collisions	AIcrowd submission
Trajectory Transformer	1.22/2.57	0.70/1.45	13.86	submission link
STGAT	0.88/1.85	0.65/1.35	11.28	submission link
Social-STGCNN	0.80/1.52	0.61/1.18	16.78	submission link
CausalHTP	0.85/1.78	0.65/1.35	10.87	submission link
SocialWays*	2.57/4.69	2.56/4.67	0.57	submission link
SR-LSTM	0.80/1.58	0.80/1.58	10.84	submission link
DAG-Net	0.66/1.44	0.56/1.22	9.22	submission link

* - unreliable official implementation, yields unreliable evaluation results

License

MIT License

Citation

If you find this code useful in your research then please cite

@article{Kothari2020HumanTF,
  author={Kothari, Parth and Kreiss, Sven and Alahi, Alexandre},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  title={Human Trajectory Forecasting in Crowds: A Deep Learning Perspective},
  year={2021},
  volume={},
  number={},
  pages={1-15},
  doi={10.1109/TITS.2021.3069362}
 }