PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection

[paper]

⭐️ Highlights

TL;DR: We propose PointOBB-v2, a simpler, faster, and stronger method to generate pseudo rotated boxes from points without relying on any other prior. It achieves a training speed 15.58x faster and an accuracy improvement of 11.60%/25.15%/21.19% on the DOTA-v1.0/v1.5/v2.0 datasets compared to the previous state-of-the-art, PointOBB.

🛠️ Installation

Please refer to the Installation, we copy it here.

conda create -n open-mmlab python=3.7 pytorch==1.7.0 cudatoolkit=10.1 torchvision -c pytorch -y conda activate open-mmlab pip install openmim mim install mmcv-full mim install mmdet git clone https://github.com/taugeren/PointOBB-v2.git cd mmrotate pip install -r requirements/build.txt pip install -v -e .

📊 Data Preparation

Please follow data_preparation to prepare formatting data

🏋️ Train CPM

If you want to visualize CPM result during training, please set visualize=True in train_config

Please modified the config code that contains the visualize directory path

The learning rate for n GPU card and batch size m is 0.0125 * n * m

For single GPU

```

Basic format: python tools/train.py ${CONFIG_FILE} [optional arguments]

python tools/train.py configs/pointobbv2/traincpmdotav10.py --work-dir workdirs/cpmdotav10 --gpu-ids 0 ```

For multiple GPU

```

Basic format: ./tools/disttrain.sh ${CONFIGFILE} ${GPU_NUM} [optional arguments]

CUDAVISIBLEDEVICES=0,1 PORT=29801 ./tools/disttrain.sh configs/pointobbv2/traincpm_dotav10.py 2 ```

🏋️ Generate Pseudo Label

Please modified the config code that contains the directory path

For single GPU

```

Basic fromat: python tools/train.py ${CONFIGFILE} --resume-from ${CPMCHECKPOINT_FILE} [other arguments]

python tools/train.py configs/pointobbv2/generatepseudolabeldotav10.py --resume-from workdirs/cpmdotav10/epoch6.pth --work-dir workdirs/cpmdotav10 --gpu-ids 0 ```

For multiple GPU

```

Basic format: ./tools/disttrainresume.sh ${CONFIGFILE} ${CPMCHECKPOINTFILE} ${GPUNUM} [optional arguments]

CUDAVISIBLEDEVICES=0,1 PORT=29801 ./tools/disttrainresume.sh /ssd1/renbotao/githubsubmission/mmrotate/configs/pointobbv2/generatepseudolabeldotav10.py workdirs/cpmdotav10/epoch_6.pth 2 ```

🏋️ Train Detector

You can use different oriented object detection detector in MMRotate Config

Please modify the pseudo label path in config file

For example, using Redet, the training command:

```

single GPU

python tools/train.py configs/pointobbv2/redetdotav10.py --work-dir workdirs/cpm_dotav10 --gpu-ids 0

multiple GPU

CUDAVISIBLEDEVICES=0,1 PORT=29801 ./tools/disttrain.sh configs/pointobbv2/redetdotav10.py 2 ```

the testing command:

```

single GPU

python tools/test.py workdirs/redetdotav10/redetdotav10.py workdirs/redetdotav10/epoch12.pth --gpu-ids 0 --format-only --eval-options submissiondir=testmodel/redetdotav10_epoch12

multiple GPU

CUDAVISIBLEDEVICES=0,1,2,3 PORT=29816 tools/disttest1.sh workdirs/redetdotav10/redetdotav10.py workdirs/redetdotav10/epoch_12.pth 4 ```

🚀 Released Models

| Dataset | Config | Log | Checkpoint | mAP | | :-------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :---: | | DOTA-v1.0 | config | quark hugging face | quark hugging face | 44.85 | | DOTA-v1.5 | config | quark hugging face | quark hugging face | 36.39 | | DOTA-v2.0 | config | quark hugging face | quark hugging face | 27.22 |

🖊️ Citation

If you find this work helpful for your research, please consider giving this repo a star ⭐ and citing our papers:

```bibtex @article{pointobbv2, title={PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection}, author={Ren, Botao and Yang, Xue and Yu, Yi and Luo, Junwei and Deng, Zhidong}, journal={arXiv preprint arXiv:2410.08210}, year={2024} }

@inproceedings{pointobb, title={PointOBB: Learning Oriented Object Detection via Single Point Supervision}, author={Luo, Junwei and Yang, Xue and Yu, Yi and Li, Qingyun and Yan, Junchi and Li, Yansheng}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={16730--16740}, year={2024} }

@inproceedings{point2rbox, title={Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision}, author={Yu, Yi and Yang, Xue and Li, Qingyun and Da, Feipeng and Dai, Jifeng and Qiao, Yu and Yan, Junchi}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={16783--16793}, year={2024} } ```

📃 License

This project is released under the Apache License 2.0.