Rep-RTADet

This is the official code repository for "Rep-RTADet: Reparameterized Real-Time Algae Object Detectors Enhanced through Dynamic Cache-Based Poisson Fusion".

Rep-RTADet 算法在阿里天池 IEEE Cybermatics 第二届国际 "Vision Meets Algae" 挑战赛 中获得冠军

环境配置

conda create -n mmyolo python=3.8 pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch -y
conda activate mmyolo
pip install openmim
mim install "mmengine>=0.6.0"
mim install "mmcv>=2.0.0rc4,<2.1.0"
mim install "mmdet>=3.0.0,<4.0.0"
git clone https://github.com/open-mmlab/mmyolo.git
cd mmyolo
# Install albumentations
pip install -r requirements/albu.txt
# Install MMYOLO
mim install -v -e .

数据集配置

以 reprtadet_l_possion.py 配置文件为例

data_root  数据根目录
train_ann_file 训练集标注文件路径（json格式）
train_data_prefix 训练集图片路径
val_ann_file 验证集标注文件路径（json格式）
val_data_prefix 验证集图片路径
test_image_info 测试集标注文件路径（json格式）
test_image 测试集图片路径

数据集文件结构

├── algae
│   ├── images
│   │   ├── train
│   │   ├── val
│   │   ├── test
│   ├── annotations
│   │   ├── instances_train.json
│   │   ├── instances_val.json
│   │   ├── instances_test.json

训练命令

python tools/train.py configs/reprtadet/reprtadet_m.py

推理命令

# val
python tools/test.py configs/reprtadet/reprtadet_m.py RepRTADet-m.pth

# test
python tools/test.py configs/reprtadet/reprtadet_m_test.py RepRTADet-m.pth

模型和结果

Model	img size	box AP0.5 val	box AP val	box AP test	TTA box AP test	预训练模型	epochs
RepRTADet-m	1280	0.934	0.723		0.7515	RTMDet-m	200
RepRTADet-m2	1280	0.933	0.722	0.7460	0.7510	RTMDet-m	200

Acknowledgements

• MMYOLO [https://github.com/open-mmlab/mmyolo/tree/main)