A PyTorch implementation and pretrained models for DNTR (DeNoising Transformer R-CNN). We present DN-FPN, a plug-in that suppresses noise generated during the fusion of FPNs. In addition, we renalvate the standard R-CNN to consist of a transformer structure, namely Trans R-CNN.(base)
Required environments:
- Linux
- Python 3.6+
- PyTorch 1.3+
- CUDA 9.2+
- GCC 5+
- MMCV
- cocoapi-aitod
Installation:
This implementation is based on MMDetection 2.24.1. Assume that your environment has satisfied the above requirements, please follow the following steps for installation.
git clone https://github.com/hoiliu-0801/DNTR.git
cd DNTR/mmdet-dntr
# Install pytorch
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
# Required packages
pip install -r requirements/build.txt
pip install yapf==0.40.0
pip install numbda
pip install timm
pip install torchprofile
python setup.py develop
# Install cocoapi
pip install pycocotools
# Install aitodcocoapi
pip install cython==0.29.36
pip install "git+https://github.com/jwwangchn/cocoapi-aitod.git#subdirectory=aitodpycocotools"
#Install mmcv
#pip install mmcv==2.0.0rc4 -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7/index.html
mim install mmcv-full==1.6.0Get Started with single GPU
Training DNTR, for example :
python tools/train.py configs/aitod-dntr/aitod_DNTR_mask.py
Testing DNTR, for example :
python tools/test.py configs/aitod-dntr/aitod_DNTR_mask.py
Table 1. Training Set: AI-TOD trainval set, Validation Set: AI-TOD test set, 36 epochs, where FRCN, DR denotes Faster R-CNN and DetectoRS, respectively.
| Method | Backbone | mAP | AP50 | AP75 | APvt | APt | APs | APm | APm |
|---|---|---|---|---|---|---|---|---|---|
| FRCN | R-50 | 11.1 | 26.3 | 7.6 | 0.0 | 7.2 | 23.3 | 33.6 | 22.9 |
| ATSS | R-50 | 12.8 | 30.6 | 8.5 | 1.9 | 11.6 | 19.5 | 29.2 | 25.8 |
| ATSS w/ DN-FPN | R-50 | 17.9 | 41.0 | 12.9 | 3.7 | 16.4 | 25.3 | 35.0 | 33.2 |
| NWD-RKA | R-50 | 23.4 | 53.5 | 16.8 | 8.7 | 23.8 | 28.5 | 36.0 | 6.9 |
| DNTR | R-50 | 26.2 | 56.7 | 20.2 | 12.8 | 26.4 | 31.0 | 37.0 | 7.6 |
Table 2. Training Set: Visdrone train set, Validation Set: Visdrone val set, 12 epochs,
| Method | Backbone | AP | AP50 | AP75 |
|---|---|---|---|---|
| DNTR | R-50 | 34.4 | 57.9 | 35.3 |
| UFPMP w/o DN-FPN | R-50 | 36.6 | 62.4 | 36.7 |
| UFPMP w/ DN-FPN | R-50 | 37.8 | 62.7 | 38.6 |
If you want to run other baseline method, please replace /mmdet/models/detectors/two_stage.py with mmdet/models/detectors/two_stage.py For example: Faster R-CNN: python tools/train.py configs/aitod-dntr/aitod_faster_r50_dntr_1x.py
@ARTICLE{10518058, author={Liu, Hou-I and Tseng, Yu-Wen and Chang, Kai-Cheng and Wang, Pin-Jyun and Shuai, Hong-Han and Cheng, Wen-Huang}, journal={IEEE Transactions on Geoscience and Remote Sensing}, title={A DeNoising FPN With Transformer R-CNN for Tiny Object Detection}, year={2024}, volume={62}, number={}, pages={1-15}, }