This repo is the top1 solution (team XXXR) for PRCV2024 SARDet contest.

• update README
• add code

challenges:

1. SARDet100k is of the most Complex(large, multi-scale, diverse sources and polarization, small objects) SAR detection datasets, which takes several times the training time of a general dataset.
2. Training strategies for optical images may not always be applicable to SAR datasets.

key ideas:

1. Train faster:
   • use real-time network YOLOv10-xl as baseline to get faster training speed
   • Introduce starnet to get a variant model
   • Cross-combining the two previously trained models to get extral two models with good pre-trained weights
2. The right data augmentations
   • larger size
   • smaller scale variation
   • weaker hue variation
3. Multi-stage training
   • smaller learning rate
   • weaker and more augmentations
4. TTA and multi-model results integration
   • TTA (flip)
   • Weighted Boxes Fusion(WBF) of 4 models

pip install -r requirements.txt
pip install -e .
pip install einops ensemble_boxes timm

# cd to code directory
mkdir datasets
# convert to YOLO format
python converter.py datasets/source/Annotations
cp -r datasets/source/JPEGImages/train datasets/SARDet2024/train/images
cp -r datasets/source/JPEGImages/val datasets/SARDet2024/val/images
# combine train val
python datapro_SARDet.py
# link test data to default dir
ln -s dir_of_test datasets/SARDet2024/test_A

# The computing resources of the organizer are 8 RTX3090
yolo detect train data=SARDet_tv.yaml model=yolov10x.pt epochs=200 batch=48 imgsz=1024 project=runs/sardet_x_1024_e200_b48 close_mosaic=0 device=0,1,2,3,4,5,6,7

python test_single.py runs/sardet_x_1024_e200_b48/train/weights/last.pt