SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving

Introduction

This is the project page of the paper

Lu Zhang, Peiliang Li, Sikang Liu, and Shaojie Shen, "SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving", arXiv preprint arXiv:2402.02519 (2024), (Corresponding author: Lu ZHANG, lzhangbz@connect.ust.hk),

which is accepted for publication in the IEEE Robotics and Automation Letters (RA-L), 2024.

Preprint: arXiv

Video: YouTube

Qualitative Results

On Argoverse 1 motion forecasting dataset

On Argoverse 2 motion forecasting dataset

Todo List

Release code for Argoverse 2 dataset
Release training and evaluation scripts for DDP
First release

Gettting Started

Install dependencies

Create a new conda virtual env

conda create --name simpl python=3.8
conda activate simpl

Install PyTorch according to your CUDA version. We recommend CUDA >= 11.1, PyTorch >= 1.8.0.

conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge

Install Argoverse 1 & 2 APIs, please follow argoverse-api and av2-api.
Install other dependencies

pip install scikit-image IPython tqdm ipdb tensorboard

Play with pretrained models (Argoverse 1)

Generate a subset of the dataset for testing using the script. It will generate 1k samples to data_argo/features/:

sh scripts/argo_preproc_small.sh

The dataset directory should be organized as follows:

data_argo
├── features
│   ├── train
│   │   ├── 100001.pkl
│   │   ├── 100144.pkl
│   │   ├── 100189.pkl
...
│   └── val
│       ├── 10018.pkl
│       ├── 10080.pkl
│       ├── 10164.pkl
...

The pre-trained weights are located at saved_models/. Use the script below to visualize prediction results:

sh scripts/simpl_av1_vis.sh

Since we store each sequence as a single file, the system may raise error OSError: [Erron 24] Too many open files during evaluation and training. You may use the command below to solve this issue:

ulimit -SHn 51200
ulimit -s unlimited

To evaluate the trained models:

sh scripts/simpl_av1_eval.sh

You are supposed to get:

Validation set finish, cost 289.01 secs
-- minade_1: 1.428 minfde_1: 3.240 mr_1: 0.512 brier_fde_1: 3.240 minade_k: 0.658 minfde_k: 0.947 mr_k: 0.081 brier_fde_k: 1.558

Train from scratch

Preprocess full Argoverse 1 motion forecasting dataset using the script:

sh scripts/argo_preproc_all.sh

The preprocessed dataset will cost about 15 GB storage, please reserve enough space for preprocessing.

Launch training using the script:

# single-GPU
sh scripts/simpl_av1_train.sh

# multi-GPU based on DDP
sh scripts/simpl_av1_train_ddp.sh

For model evaluation, please refer to the following scripts:

# single-GPU
sh scripts/simpl_av1_eval.sh

# multi-GPU based on DDP
sh scripts/simpl_av1_eval_ddp.sh

Train/Evaluate/Visualize SIMPL using Argoverse 2

Please refer to the scripts in the directory scripts/, and the usage is similar to scripts for the Argoverse 1 dataset. If you have any questions, please feel free to raise an issue or contact us via email.

Acknowledgment

We would like to express sincere thanks to the authors of the following packages and tools:

License

This repository is licensed under MIT license.

HKUST-Aerial-Robotics/SIMPL