Panacea: Panoramic and Controllable Video Generation for Autonomous Driving

Official Repository of Panacea.

[Paper] Panacea: Panoramic and Controllable Video Generation for Autonomous Driving,
Yuqing Wen^1*†, Yucheng Zhao^2*,Yingfei Liu^2*, Fan Jia², Yanhui Wang¹, Chong Luo¹, Chi Zhang³, Tiancai Wang^2‡, Xiaoyan Sun^1‡, Xiangyu Zhang^2
1University of Science and Technology of China, ²MEGVII Technology, ³Mach Drive
^*Equal Contribution, ^†This work was done during the internship at MEGVII, ^‡Corresponding Author.

[Paper] Panacea+: Panoramic and Controllable Video Generation for Autonomous Driving,
Yuqing Wen^1*†, Yucheng Zhao^2*,Yingfei Liu^2*, Binyuan Huang^4*, Fan Jia², Yanhui Wang¹, Chi Zhang³, Tiancai Wang^2‡, Xiaoyan Sun^1‡, Xiangyu Zhang^2
1University of Science and Technology of China, ²MEGVII Technology, ³Mach Drive, ⁴Wuhan University
^*Equal Contribution, ^†This work was done during the internship at MEGVII, ^‡Corresponding Author.

[WebPage] https://panacea-ad.github.io/

News

• Aug. 15th, 2024: We release an enhanced version of Panacea, named Panancea+, which has improved performance and comprehensive validation on multiple datasets and tasks. For more details, please refer to the paper Panacea+.

• Aug. 15th, 2024: We release the checkpoint and inference scripts for stage 2 of Panacea+, you can use it to generate multi-view video samples based on BEV layout sequences.

• Apr. 18th, 2024: We release our Gen-nuScenes dataset generated by Panacea. Please check the metrics/ folder to use it.

• Apr. 18th, 2024: We release the BEV-perception evaluation codes based on StreamPETR. Please check the metrics/ folder and follow the metrics/README.md for detailed evaluation.

Getting Started

Please follow our documentation step by step.

Environment Setup

Following the instruction from: Environment Setup.

Prepare dataset

Prepare real dataset following the instruction from Data Preparation.

Remember to put the dataset under the path data/nuscenes

Download pretrained checkpoint

Download the weights of the second stage from panaceaplus_40k_deepspeed.ckpt

Put it to folder checkpoints/

Inference

--split: to specify train or val sets

--use_last_frame=true means use the last frame as conditional image.

Run the following command to inference stage 2 on the whole training/val set of nuscenes.

python -m torch.distributed.launch --nproc_per_node=8 --master_port=1238 inference.py --base configs/inference_nuscenes.yaml --ckptpath --ckpt checkpoints/panaceaplus_40k_deepspeed.ckpt --split train --use_last_frame true --name EXP_NAME --bs 1

Generating Multi-View and Controllable Videos for Autonoumous Driving

Overview of Panacea. (a). The diffusion training process of Panacea, enabled by a diffusion encoder and decoder with the decomposed 4D attention module. (b). The decomposed 4D attention module comprises three components: intra-view attention for spatial processing within individual views, cross-view attention to engage with adjacent views, and cross-frame attention for temporal processing. (c). Controllable module for the integration of diverse signals. The image conditions are derived from a frozen VAE encoder and combined with diffused noises. The text prompts are processed through a frozen CLIP encoder, while BEV sequences are handled via ControlNet. (d). The details of BEV layout sequences, including projected bounding boxes, object depths, road maps and camera pose.

The two-stage inference pipeline of Panacea. Its two-stage process begins by creating multi-view images with BEV layouts, followed by using these images, along with subsequent BEV layouts, to facilitate the generation of following frames.

🎬   BEV-guided Video Generation   🎬

Controllable multi-view video generation. Panacea is able to generate realistic, controllable videos with good temporal and view consistensy.

🎞   Attribute Controllable Video Generation   🎞

Video generation with variable attribute controls, such as weather, time, and scene, which allows Panacea to simulate a variety of rare driving scenarios, including extreme weather conditions such as rain and snow, thereby greatly enhancing the diversity of the data.

🔥   Benefiting Autonomous Driving   🔥

(a). Panoramic video generation based on BEV (Bird’s-Eye-View) layout sequence facilitates the establishment of a synthetic video dataset, which enhances perceptual tasks. (b). Producing panoramic videos with conditional images and BEV layouts can effectively elevate image-only datasets to video datasets, thus enabling the advancement of video-based perception techniques.

BibTex

                
@inproceedings{wen2024panacea,
  title={Panacea: Panoramic and controllable video generation for autonomous driving},
  author={Wen, Yuqing and Zhao, Yucheng and Liu, Yingfei and Jia, Fan and Wang, Yanhui and Luo, Chong and Zhang, Chi and Wang, Tiancai and Sun, Xiaoyan and Zhang, Xiangyu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6902--6912},
  year={2024}
}
@misc{wen2024panaceapanoramiccontrollablevideo,
      title={Panacea+: Panoramic and Controllable Video Generation for Autonomous Driving}, 
      author={Yuqing Wen and Yucheng Zhao and Yingfei Liu and Binyuan Huang and Fan Jia and Yanhui Wang and Chi Zhang and Tiancai Wang and Xiaoyan Sun and Xiangyu Zhang},
      year={2024},
      eprint={2408.07605},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.07605}, 
}
}

Contact

Feel free to contact us at wenyuqing AT mail.ustc.edu.cn or wangtiancai AT megvii.com

Acknowledgement

This code builds on Stability-AI, ControlNet and StreamPETR. Thanks for open-sourcing!