OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving
Wenzhao Zheng*, Weiliang Chen*, Yuanhui Huang, Borui Zhang, Yueqi Duan, Jiwen Lu
* Equal contribution
- [2024/3/13] We release the code for visualization, our training log and our pretrianed model.
- [2023/12/7] We update the code and config files for OccWorld.
OccWorld models the joint evolutions of 3D scenes and ego movements.
Combined with self-supervised (SelfOcc), LiDAR-collected (TPVFormer), or machine-annotated (SurroundOcc) 3D occupancy, OccWorld has the potential to scale up to large-scale training, paving the way for interpretable end-to-end large driving models.
Given past 3D occupancy observations, our self-supervised OccWorld trained can forecast future scene evolutions and ego movements jointly. This task requires a spatial understanding of the 3D scene and temporal modeling of how driving scenarios develop. We observe that OccWorld can successfully forecast the movements of surrounding agents and future map elements such as drivable areas. OccWorld even generates more reasonable drivable areas than the ground truth, demonstrating its ability to understand the scene rather than memorizing training data. Still, it fails to forecast new vehicles entering the sight, which is difficult given their absence in the inputs.
-
Create conda environment with python version 3.8.0
-
Install all the packages in environment.yaml
-
Anything about the installation of mmdetection3d, please refer to mmdetection3d
-
Create soft link from data/nuscenes to your_nuscenes_path
-
Prepare the gts semantic occupancy introduced in Occ3d
-
Download our generated train/val pickle files and put them in data/
The dataset should be organized as follows:
OccWorld/data
nuscenes - downloaded from www.nuscenes.org
lidarseg
maps
samples
sweeps
v1.0-trainval
gts - download from Occ3d
nuscenes_infos_train_temporal_v3_scene.pkl
nuscenes_infos_val_temporal_v3_scene.pkl
- We also provide our pretrained model in https://cloud.tsinghua.edu.cn/d/ff4612b2453841fba7a5/ .
Train the VQVAE on RTX 4090 with 24G GPU memory.
python train.py --py-config config/train_vqvae.py --work-dir out/vqvae
Train the OccWorld on RTX 4090 with 24G GPU memory. (Remember to change the checkpoint path of VQVAE in the config file)
python train.py --py-config config/train_occworld.py --work-dir out/occworld
Eval the model on RTX 4090 with 24G GPU memory. (Remember to change the checkpoint path of OccWorld in the config file)
python eval_metric_stp3.py --py-config config/occworld.py --work-dir out/occworld
Visualize the results use the following code.
python visualize_demo.py --py-config config/train_occworld.py --work-dir out/occworld
Also, you can specific the visulized scene index by adding --scene-idx i0 i1 i2 ...
. After running the above code, two folders will
be created: "i_input" and "i." The "i" folder will contain the autoregressive predicted results
for t=0.5, 1, 1.5, 2, 2.5, and 3. The "i_input" folder will contain visualized ground truth results for t=0, 0.5, 1, 1.5, 2, 2.5, and 3.
Our code is based on TPVFormer, SelfOcc, and PointOcc.
Also thanks to these excellent open-sourced repos: SurroundOcc OccFormer BEVFormer
If you find this project helpful, please consider citing the following paper:
@article{zheng2023occworld,
title={OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving},
author={Zheng, Wenzhao and Chen, Weiliang and Huang, Yuanhui and Zhang, Borui and Duan, Yueqi and Lu, Jiwen },
journal={arXiv preprint arXiv: 2311.16038},
year={2023}
}