Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction

This repo is the official implementation of "Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction" by Zhuofan Zong, Dongzhi Jiang, Guanglu Song, Zeyue Xue, Jingyong Su, Hongsheng Li, and Yu Liu.

News

• [07/25/2023] Code for HoP on BEVDet is released!
• [07/14/2023] HoP is accepted to ICCV 2023!
• [04/05/2023] HoP achieves new SOTA performance on nuScenes 3D detection leaderboard with 68.5 NDS and 62.4 mAP.

Model Zoo

Result on BEVDet4D-Depth

| model | backbone | pretrain | img size | Epoch | NDS | mAP | config | ckpt | log | | :----------------------: | :------: | :----------: | :------: | :---: | :----: | :----: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | | BEVDet4D-Depth(Baseline) | Res50 | ImageNet | 256x704 | 24 | 0.4930 | 0.3848 | cfg | ckpt | log | | HoP_BEVDet4D-Depth | Res50 | ImageNet | 256x704 | 24 | 0.5099 | 0.3990 | cfg | ckpt | log |

Get Started

Install

We train our models under the following environment:

python=3.6.9 pytorch=1.8.1 torchvision=0.9.1 cuda=11.2

Other versions may possibly be imcompatible.

We use MMDetection3D V1.0.0rc4, MMDetection V2.24.0 and MMCV V1.5.0. The source code of MMDetection3D has been included in this repo.

You can take the following steps to install packages above:

1. Build MMCV following official instructions.

2. Install MMDetection by

bash pip install mmdet==2.24.0

1. Copy HoP repo and install MMDetection3D.

bash git clone git@github.com:Sense-X/HoP.git cd HoP pip install -e .

Data Preparation

Follow the steps to prepare nuScenes Dataset introduced in nuscenes_det.md and create the pkl by running:

bash python tools/create_data_bevdet.py

Train HoP

```bash

single gpu

python tools/train.py configs/hopbevdet/hopbevdet4d-r50-depth.py

multiple gpu

./tools/disttrain.sh configs/hopbevdet/hopbevdet4d-r50-depth.py $numgpu ```

Eval HoP

```bash

single gpu

python tools/test.py configs/hopbevdet/hopbevdet4d-r50-depth.py $checkpoint --eval bbox

multiple gpu

./tools/disttest.sh configs/hopbevdet/hopbevdet4d-r50-depth.py $checkpoint $numgpu --eval bbox ```

Method

TODO

• [ ] Release code for HoP on BEVFormer.

Cite HoP

If you find this repository useful, please use the following BibTeX entry for citation.

latex @misc{hop2023, title={Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction}, author={Zhuofan Zong and Dongzhi Jiang and Guanglu Song and Zeyue Xue and Jingyong Su and Hongsheng Li and Yu Liu}, year={2023}, eprint={2304.00967}, archivePrefix={arXiv}, primaryClass={cs.CV} }

License

This project is released under the MIT license. Please see the LICENSE file for more information.