ImageDataBunch from 500 megapixel images as tiles

fastai
25

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
    path = path_lbl / path.name
    return ImageTile(path, *tile)


def get_tiles(images: PathOrStr, rows: int, cols: int) -> Collection[ImageTile]:
    images_tiles = []
    for img in images:
        for i in range(rows * cols):
            images_tiles.append(ImageTile(img, i, rows, cols))
    return images_tiles


def open_image_tile(img_t: ImageTile, mask=False, **kwargs) -> Image:
    """given and ImageTile it returns and Image with the tile,
    set mask to True for masks"""
    path, idx, rows, cols = img_t
    img = open_image(path, **kwargs) if not mask else open_mask(path, **kwargs)
    row = idx // cols
    col = idx % cols
    tile_x = img.size[0] // cols
    tile_y = img.size[1] // rows
    return Image(img.data[:, col * tile_x:(col + 1) * tile_x, row * tile_y:(row + 1) * tile_y])


class SegmentationTileLabelList(SegmentationLabelList):

    def open(self, fn: ImageTile):
        return open_image_tile(fn, div=True, mask=True)


class SegmentationTileItemList(ImageList):
    _label_cls, _square_show_res = SegmentationTileLabelList, False

   
    def open(self, fn: ImageTile) -> Image:
        return open_image_tile(fn, convert_mode=self.convert_mode, after_open=self.after_open)

    @classmethod
    def from_folder(cls, path: PathOrStr = '.', rows=1, cols=1, extensions: Collection[str] = None, **kwargs) -> ItemList:
        """patchs the from_folder method, generating list of ImageTile with all the possible tiles for all the images in folder"""
        files = get_files(path, extensions, recurse=True)
        files_tiled = get_tiles(files, rows, cols)
        return SegmentationTileItemList(files_tiled, **kwargs)

Hope this is not a silly approach and it can bu useful.

3 Likes