/ST-P3

[ECCV 2022] ST-P3, an end-to-end vision-based autonomous driving framework via spatial-temporal feature learning.

Primary LanguagePythonApache License 2.0Apache-2.0

ST-P3

pipeline

ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning
Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, Dacheng Tao.

Introduction

This reposity is the official PyTorch Lightning implementation for ST-P3.

TL;DR: we propose a spatial-temporal feature learning scheme towards a set of more representative features for perception, prediction and planning tasks simultaneously in autonomous driving, and thus devise an explicit pipeline to generate planning trajectories directly from raw sensor inputs.

Get Started

Setup

conda env create -f environment.yml
git clone https://github.com/OpenPerceptionX/ST-P3.git

Pre-trained models

  • open-loop planning on nuScenes: model.
  • closed-loop planning on CARLA: model.

Evaluation

To evaluate the model on nuScenes:

  • Download the nuScenes dataset.
  • Download the pretrained weights.
bash scripts/eval_plan.sh ${checkpoint} ${dataroot}

To evaluate the model on CARLA:

  • Please refer to the Transfuser to set up the environment.
  • Test with the carla_agent.py file and the pretrained weights.

Training

# (recommended) perception module pretrain
bash scripts/train_perceive.sh ${configs} ${dataroot}

# (optional) prediction module training purpose, no need for e2e training
bash scripts/train_prediction.sh ${configs} ${dataroot} ${pretrained}

# entire model e2e training
bash scripts/train_plan.sh ${configs} ${dataroot} ${pretrained}
  • To train the model from scratch on nuScenes, we recommend to train a perceptual weight first and use it to train subsequent tasks to prevent nan during training.
  • If you would like to use the nuScenes depth data (will be released very soon), put the depth folder in the dataroot directory and change GT_DEPTH in the config file to True.

Benchmark

  • Open-loop planning results on nuScenes.
Method L2 (m) 1s L2 (m) 2s L2 (m) 3s Collision (%) 1s Collision (%) 2s Collision (%) 3s
Vanilla 0.50 1.25 2.80 0.68 0.98 2.76
NMP 0.61 1.44 3.18 0.66 0.90 2.34
Freespace 0.56 1.27 3.08 0.65 0.86 1.64
ST-P3 1.33 2.11 2.90 0.23 0.62 1.27
  • Closed-loop simulation results on CARLA.
Method Town05 Short DS Town05 Short RC Town05 Long DS Tow05 Long RC
CILRS 7.47 13.40 3.68 7.19
LBC 30.97 55.01 7.05 32.09
Transfuser 54.52 78.41 33.15 56.36
ST-P3 55.14 86.74 11.45 83.15

Visualization

  • nuScenes visualization results

nuscenes_vis

  • CARLA visualization results

CARLA_vis

Citation

If you find our repo or our paper useful, please use the following citation:

@inproceedings{hu2022stp3,
 title={ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning}, 
 author={Shengchao Hu and Li Chen and Penghao Wu and Hongyang Li and Junchi Yan and Dacheng Tao},
 booktitle={European Conference on Computer Vision (ECCV)},
 year={2022}
}

License

All code within this repository is under Apache License 2.0.

Acknowledgement

We thank Xiangwei Geng for his support on the depth map generation, and fruitful discussions from Xiaosong Jia. We have many thanks to FIERY team for their exellent open source project.