BEVDepth is a new 3D object detector with a trustworthy depth estimation. For more details, please refer to our paper on Arxiv.
BEVStereo is a new multi-view 3D object detector using temporal stereo to enhance depth estimation.
MatrixVT is a novel View Transformer for BEV paradigm with high efficiency and without customized operators. For more details, please refer to our paper on Arxiv. Try MatrixVT on CPU by run this file !
- γ2022/12/06γ We released our new View Transformer (MatrixVT), the paper is on Arxiv.
- γ2022/11/30γ We updated our paper(BEVDepth) on Arxiv.
- γ2022/11/18γ Both BEVDepth and BEVStereo were accepted by AAAI'2023.
- γ2022/09/22γ We released our paper(BEVStereo) on Arxiv.
- γ2022/08/24γ We submitted our result(BEVStereo) on nuScenes Detection Task and achieved the SOTA.
- γ2022/06/23γ We submitted our result(BEVDepth) without extra data on nuScenes Detection Task and achieved the SOTA.
- γ2022/06/21γ We released our paper(BEVDepth) on Arxiv.
- γ2022/04/11γ We submitted our result(BEVDepth) on nuScenes Detection Task and achieved the SOTA.
Step 0. Install pytorch(v1.9.0).
Step 1. Install MMDetection3D(v1.0.0rc4).
Step 2. Install requirements.
pip install -r requirements.txt
Step 3. Install BEVDepth(gpu required).
python setup.py develop
Step 0. Download nuScenes official dataset.
Step 1. Symlink the dataset root to ./data/
.
ln -s [nuscenes root] ./data/
The directory will be as follows.
BEVDepth
βββ data
β βββ nuScenes
β β βββ maps
β β βββ samples
β β βββ sweeps
β β βββ v1.0-test
| | βββ v1.0-trainval
Step 2. Prepare infos.
python scripts/gen_info.py
Train.
python [EXP_PATH] --amp_backend native -b 8 --gpus 8
Eval.
python [EXP_PATH] --ckpt_path [CKPT_PATH] -e -b 8 --gpus 8
Exp | EMA | CBGS | mAP | mATE | mASE | mAOE | mAVE | mAAE | NDS | weights |
---|---|---|---|---|---|---|---|---|---|---|
BEVDepth | 0.3304 | 0.7021 | 0.2795 | 0.5346 | 0.5530 | 0.2274 | 0.4355 | github | ||
BEVDepth | β | 0.3329 | 0.6832 | 0.2761 | 0.5446 | 0.5258 | 0.2259 | 0.4409 | github | |
BEVDepth | β | 0.3484 | 0.6159 | 0.2716 | 0.4144 | 0.4402 | 0.1954 | 0.4805 | github | |
BEVDepth | β | β | 0.3589 | 0.6119 | 0.2692 | 0.5074 | 0.4086 | 0.2009 | 0.4797 | github |
BEVStereo | 0.3456 | 0.6589 | 0.2774 | 0.5500 | 0.4980 | 0.2278 | 0.4516 | github | ||
BEVStereo | β | 0.3494 | 0.6671 | 0.2785 | 0.5606 | 0.4686 | 0.2295 | 0.4543 | github | |
BEVStereo | 0.3427 | 0.6560 | 0.2784 | 0.5982 | 0.5347 | 0.2228 | 0.4423 | github | ||
BEVStereo | β | 0.3435 | 0.6585 | 0.2757 | 0.5792 | 0.5034 | 0.2163 | 0.4485 | github | |
BEVStereo | β | 0.3576 | 0.6071 | 0.2684 | 0.4157 | 0.3928 | 0.2021 | 0.4902 | github | |
BEVStereo | β | β | 0.3721 | 0.5980 | 0.2701 | 0.4381 | 0.3672 | 0.1898 | 0.4997 | github |
- The results are different between evaluation during training and evaluation from ckpt.
Due to the working mechanism of EMA, the model parameters saved by ckpt are different from the model parameters used in the training stage.
- EMA exps are unable to resume training from ckpt.
We used the customized EMA callback and this function is not supported for now.
If you use BEVDepth and BEVStereo in your research, please cite our work by using the following BibTeX entry:
@article{li2022bevdepth,
title={BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object Detection},
author={Li, Yinhao and Ge, Zheng and Yu, Guanyi and Yang, Jinrong and Wang, Zengran and Shi, Yukang and Sun, Jianjian and Li, Zeming},
journal={arXiv preprint arXiv:2206.10092},
year={2022}
}
@article{li2022bevstereo,
title={Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo},
author={Li, Yinhao and Bao, Han and Ge, Zheng and Yang, Jinrong and Sun, Jianjian and Li, Zeming},
journal={arXiv preprint arXiv:2209.10248},
year={2022}
}
@article{zhou2022matrixvt,
title={MatrixVT: Efficient Multi-Camera to BEV Transformation for 3D Perception},
author={Zhou, Hongyu and Ge, Zheng and Li, Zeming and Zhang, Xiangyu},
journal={arXiv preprint arXiv:2211.10593},
year={2022}
}