This repository contains the code for the paper

MG-GAN: A Multi-Generator Model Preventing Out-of-Distribution Samples in Pedestrian Trajectory Prediction
Patrick Dendorfer*, Sven Elflein*, Laura Leal-Taixé (* equal contribution)
International Conference on Computer Vision (ICCV), 2021

The distribution over future trajectories of pedestrians is often multi-modal and does not have connected support (a).

We found that single generator GANs introduce out-of-distribution (OOD) samples in this case due to GANs mapping the continuous latent variable z with a continuous function (b). These OOD samples might introduce unforseen behavior in real world applications, such as autonomous driving.

To resolve this problem, we propose to learn the target distribution in a piecewise manner using multiple generators, effectively preventing OOD samples (c).

Our model consists of four key components: Encoding modules, Attention modules, and our novel contribution PM-Network learning a distribution over multiple Generators.

First, setup Python environment

conda create -f environment.yml -n mggan
conda activate mggan

Then, download the datasets (data.zip) from here and unzip in the root of this repository

unzip data.zip

which will create a folder ./data/datasets.

Models can be trained using the script mggan/model/train.py using the following command

python mggan/models/pinet_multi_generator/train.py --name <name_of_experiment> --num_gens <number_of_generators>  --dataset <dataset_name> --epochs 50

This generates a output folder in ./logs/<name_of_experiment> with Tensorboard logs and the model checkpoints. You can use tensorboard --logdir ./logs/<name_of_experiment> to monitor the training process.

For evaluation of metrics (ADE, FDE, Precison, Recall) for k=1 to k=20 predictions, use

python scripts/evaluate.py --model_path <path_to_model_directory>  --output_folder <folder_to_store_result_csv>

One can use --eval-set <dataset_name> to evaluate models on other test sets than the dataset the model was trained on. This is useful to evaluate the BIWI models on the Garden of Forking Paths dataset (gofp) for which we report results in the paper.

We provide pre-trained models for MG-GAN with 2-8 generators together with the training configurations, on the BIWI datasets and Stanford Drone dataset (SDD) here.

If our work is useful to you, please consider citing

@inproceedings{dendorfer2021iccv,
  title={MG-GAN: A Multi-Generator Model Preventing Out-of-Distribution Samples in Pedestrian Trajectory Prediction}, 
  author={Dendorfer, Patrick and Elflein, Sven and Leal-Taixé, Laura},
  month={October}
  year={2021},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  }