Intuition about approach to a contest

[s.s.o]

Two approaches may work - which one is better I have no idea.

for tiling to mask you can check:

ImageDataBunch from 500 megapixel images as tiles

I used this solution. I created a custom open() method that takes a tuple of file path and tile index and returns a fastai Image of the selected tile. Then I customized also from_folder() method. ImageTile = namedtuple('ImageTile', 'path idx rows cols') def calc_n_tiles(size, tile_max_size): x, y = size n_cols = x % (x // tile_max_size + 1) n_rows = y % (y // tile_max_size + 1) return n_rows, n_cols, (x//n_cols, y//n_rows) def get_labels_tiles(fn): path, *tile = fn p…

there are also treads about satellite imaging and kaggle notebooks.
sequence of images to mask you can check:

Image sequences – fastai