Draft plan

1. Decode masks to rotated bounding boxes

a. From mark rle to rotated bounding boxes

b. TO be discussed: is this a good approach? Main reason, less outputs to learn and it enabled me to use lesson 8 and 9 from fastAi course.

2. Learn coordinates 1 ship

a. Create a CSV containing all pictures including 1 ship, and add rotated bboxes to it

b. Updated CoordTransform to enable augmentation for these bounding boxes

c. Learn with fastai 'default' settings

3. Implement lesson 9 for airbus dataset

And first submit to kaggle

4. Try and implement other ideas from lesson 10 onwards