jacobieee / st-mr

Graph-based Spatial Transformer with Memory Replay for Multi-future Pedestrian Trajectory Prediction

ST-MR

Graph-based Spatial Transformer with Memory Replay for Multi-future Pedestrian Trajectory Prediction.

This paper has been accepted by CVPR 2022.

If you find this work useful for your research, please cite:

@inproceedings{li2022graph,
  title={Graph-Based Spatial Transformer With Memory Replay for Multi-Future Pedestrian Trajectory Prediction},
  author={Li, Lihuan and Pagnucco, Maurice and Song, Yang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2231--2241},
  year={2022}
}

Dataset

The dataset we use is VIRAT/ActEV and The Forking Paths.

Preprocessing.

VIRAT/ActEV.

Follow the preprocessing step in Multiverse repo and generate three .npz files with all trajectory inputs and features. Note that this step can occupy large memory. It's recommended to run it on a machine with at least 16g RAM. Or you can prepare the dataset separately.

The Forking Paths.

Follow here to download and prepare the dataset and visualize it.

Training

Requirements are listed in requirements.txt.

python code/train.py actev_preprocess models my_model/ \
--wd 0.001 --runId 0 --obs_len 8 --pred_len 12 --emb_size 32 \
--enc_hidden_size 256 --dec_hidden_size 256 --activation_func tanh \
--keep_prob 1.0 --num_epochs 80 --batch_size 20 --init_lr 0.3 --use_gnn \
--use_scene --learning_rate_decay 0.95 --num_epoch_per_decay 2.0 \
--grid_loss_weight 1.0 --grid_reg_loss_weight 0.2 --save_period 2000 \
--scene_h 36 --scene_w 64 --scene_conv_kernel 3 --scene_conv_dim 64 \
--scene_grid_strides 2,4 --use_grids 1,1 --val_grid_num 0 --train_w_onehot --gpuid 0 \
--add_self_attn --add_mr

Testing single-future prediction

python code/test.py actev_preprocess models my_model/ \
--wd 0.001 --runId 0 --obs_len 8 --pred_len 12 --emb_size 32 \
--enc_hidden_size 256 --dec_hidden_size 256 --activation_func tanh \
--keep_prob 1.0 --num_epochs 80 --batch_size 20 --init_lr 0.3 --use_gnn \
--use_scene --learning_rate_decay 0.95 --num_epoch_per_decay 2.0 \
--grid_loss_weight 1.0 --grid_reg_loss_weight 0.2 --save_period 2000 \
--scene_h 36 --scene_w 64 --scene_conv_kernel 3 --scene_conv_dim 64 \
--scene_grid_strides 2,4 --use_grids 1,0 --val_grid_num 0 --gpuid 0 \
--add_self_attn --add_mr --load_best

Testing multi-future prediction

Run inference and visualization, please refer to here.

run minADEk and minFDEk, as well as generate a trajectory usage file named "traj_usage". This is a pickles file that contains the PTU result.

python code/multifuture_eval_trajs.py forking_paths_dataset/next_x_v1_dataset_prepared_data/multifuture/test/ \
my_traj.traj.p traj_usage

Run Negative Log-likelihood.

python code/multifuture_eval_trajs_prob.py forking_paths_dataset/next_x_v1_dataset_prepared_data/multifuture/test/ \
my_prob.prob.p

For visualization, please refer to here.

Images in folder "imgs" show a set of comparison among our model, Multiverse, Social GAN and STGAT.

Run multi-future inference on sgan and stgat.

Train the models of sgan and stgat and replace the files in these two repos with those in multi_traj_sgan and multi_traj_stgat. For example, if run such inference on stgat,

1. unzip the forking_paths.zip under the corresponding folder and put it in the "dataset" folder in stgat repo. This file contains the observed trajectories.
2. Run evaluate.py:python evaluate.py. It will create a file stgat_out.npy which contains the output trajectories and files.npy which contains the name of all data sample in the Forking Paths dataset.
3. Run python get_final_output.py to process the output with the same format as my_traj.traj.p. It will create a file stgat_output.traj.p.
4. Run python code/multifuture_eval_trajs.py forking_paths_dataset/next_x_v1_dataset_prepared_data/multifuture/test/ \ stgat_output_vp.traj.p to get the result. The later steps are the same as the above sections.

Acknowledgements

Code in this repo is largely borrowed from Multiverse.