2nd ship detection challenge - Huggingface

Training a yolov8 on ship detection from aerial images.

You can find data here

Installation

pip install -r requirements.txt

Training

It is an example, you must adapt path for your use case. plz read help in the main if you don't want to preprocess data each time

#For preparing data and training yolov8
python3 main.py --origdata pathtoyourproject/shipdetection/data/original/train/ --destsplit /tmp/split --destdatayolo /tmp/yolodest --withdraw 3 --maketile False --tilesize 640 --yolov8path yolov8s --imgsz 640

Inference

To run inference with ensembling, you must run the script at pathtoprohect/shipdetection/models/yolov5_obb-master/sub_classifandyolo.py

I have adapt the model from https://github.com/hukaixuan19970627/yolov5_obb where yolov5 makes rotated bounding box. In the challenge I need rectangular bounding box, so I had to modify the code for prediction.

This part is for final solution proposed: ensembling yolov8s trained on 640x640 with yolov5 pretrained on dota.

#For preparing data and training yolov8
cd pathtoyourproject/shipdetection/models/yolov5_obb-master/

PYTHONPATH=pathtoyourproject/shipdetection/models/yolov5_obb-master/ python3 sub_classifandyolo.py

Thank you

A special thank to https://github.com/hukaixuan19970627/yolov5_obb for his pretrained model on dota

License

MIT