Position Embedding Transformation for Multi-View 3D Object Detection

This repository is an official implementation of PETR and PETRv2. The flash attention version can be find from the "flash" branch.

PETR develops position embedding transformation (PETR) for multi-view 3D object detection. PETR encodes the position information of 3D coordinates into image features, producing the 3D position-aware features. Object query can perceive the 3D position-aware features and perform end-to-end object detection. It can serve as a simple yet strong baseline for future research.

PETRv2 is a unified framework for 3D perception from multi-view images. Based on PETR, PETRv2 explores the effectiveness of temporal modeling, which utilizes the temporal information of previous frames to boost 3D object detection. The 3D PE achieves the temporal alignment on object position of different frames. A feature-guided position encoder is further introduced to improve the data adaptability of 3D PE. To support for high-quality BEV segmentation, PETRv2 provides a simply yet effective solution by adding a set of segmentation queries. Each segmentation query is responsible for segmenting one specific patch of BEV map. PETRv2 achieves state-of-the-art performance on 3D object detection and BEV segmentation.

News

2023.01.25 Our multi-view 3D detection framework StreamPETR (63.6% NDS and 55.0% mAP)** without TTA and future frames.
2023.01.04 Our multi-modal detection framework CMT is released on arxiv.
2022.11.04 The code of multi-scale improvement in PETRv2 is released.
2022.09.21 The code of query denoise improvement in PETRv2 is released.
2022.09.04 PETRv2 with VoVNet backbone and multi-scale achieves (59.1% NDS and 50.8% mAP).
2022.08.11 PETRv2 with GLOM-like backbone and query denoise achieves (59.2% NDS and 51.2% mAP) without extra data.
2022.07.04 PETR has been accepted by ECCV 2022.
2022.06.28 The code of BEV Segmentation in PETRv2 is released.
2022.06.16 The code of 3D object detection in PETRv2 is released.
2022.06.10 The code of PETR is released.
2022.06.06 PETRv2 is released on arxiv.
2022.06.01 PETRv2 achieves another SOTA performance on nuScenes dataset (58.2% NDS and 49.0% mAP) by the temporal modeling and supports BEV segmentation.
2022.03.10 PETR is released on arxiv.
2022.03.08 PETR achieves SOTA performance (50.4% NDS and 44.1% mAP) on standard nuScenes dataset.

Preparation

This implementation is built upon detr3d, and can be constructed as the install.md.

• Environments
  Linux, Python==3.6.8, CUDA == 11.2, pytorch == 1.9.0, mmdet3d == 0.17.1

• Detection Data
  Follow the mmdet3d to process the nuScenes dataset (https://github.com/open-mmlab/mmdetection3d/blob/master/docs/en/data_preparation.md).

• Segmentation Data
  Download Map expansion from nuScenes dataset (https://www.nuscenes.org/nuscenes#download). Extract the contents (folders basemap, expansion and prediction) to your nuScenes maps folder.
  Then build Segmentation dataset:

  cd tools
  python build-dataset.py
  

  If you want to train the segmentation task immediately, we privided the processed data ( HDmaps-final.tar ) at gdrive. The processed info files of segmentation can also be find at gdrive.

• Pretrained weights
  To verify the performance on the val set, we provide the pretrained V2-99 weights. The V2-99 is pretrained on DDAD15M (weights) and further trained on nuScenes train set with FCOS3D. For the results on test set in the paper, we use the DD3D pretrained weights. The ImageNet pretrained weights of other backbone can be found here. Please put the pretrained weights into ./ckpts/.

• After preparation, you will be able to see the following directory structure:

  PETR
  ├── mmdetection3d
  ├── projects
  │   ├── configs
  │   ├── mmdet3d_plugin
  ├── tools
  ├── data
  │   ├── nuscenes
  │     ├── HDmaps-nocover
  │     ├── ...
  ├── ckpts
  ├── README.md
  

Train & inference

cd PETR

You can train the model following:

tools/dist_train.sh projects/configs/petr/petr_r50dcn_gridmask_p4.py 8 --work-dir work_dirs/petr_r50dcn_gridmask_p4/

You can evaluate the model following:

tools/dist_test.sh projects/configs/petr/petr_r50dcn_gridmask_p4.py work_dirs/petr_r50dcn_gridmask_p4/latest.pth 8 --eval bbox

Visualize

You can generate the reault json following:

./tools/dist_test.sh projects/configs/petr/petr_vovnet_gridmask_p4_800x320.py work_dirs/petr_vovnet_gridmask_p4_800x320/latest.pth 8 --out work_dirs/pp-nus/results_eval.pkl --format-only --eval-options 'jsonfile_prefix=work_dirs/pp-nus/results_eval'

You can visualize the 3D object detection following:

python3 tools/visualize.py

Main Results

PETR: We provide some results on nuScenes val set with pretrained models. These model are trained on 8x 2080ti without cbgs. Note that the models and logs are also available at Baidu Netdisk with code petr.

PETRv2: We provide a 3D object detection baseline and a BEV segmentation baseline with two frames. The model is trained on 8x 2080ti without cbgs. The processed info files contain 30 previous frames, whose transformation matrix is aligned with the current frame. The info files, models and logs are also available at Baidu Netdisk with code petr.

StreamPETR: Stream-PETR achieves significant performance improvements without introducing extra computation cost, compared to the single-frame baseline.

Acknowledgement

Many thanks to the authors of mmdetection3d and detr3d .

Citation

If you find this project useful for your research, please consider citing:

@article{liu2022petr,
  title={Petr: Position embedding transformation for multi-view 3d object detection},
  author={Liu, Yingfei and Wang, Tiancai and Zhang, Xiangyu and Sun, Jian},
  journal={arXiv preprint arXiv:2203.05625},
  year={2022}
}

@article{liu2022petrv2,
  title={PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images},
  author={Liu, Yingfei and Yan, Junjie and Jia, Fan and Li, Shuailin and Gao, Qi and Wang, Tiancai and Zhang, Xiangyu and Sun, Jian},
  journal={arXiv preprint arXiv:2206.01256},
  year={2022}
}

Contact

If you have any questions, feel free to open an issue or contact us at [email protected], [email protected] or [email protected].