Multiple Band Tif

[mthsdiniz-usp]

Hey @jaspersiebring , first of all, congrats on making this library!

I'm trying to convert a set o 18 band TIF files and theirs respective geojson labels but i'm facing some problems when trying to do that.

Each one of my images has 18 bands but their height and width vary a little. When trying to feed the 18 band tif, i end up with the following error:

When trying to feed a 3 band file, i end up with the following error:

This is the code i'm using:

# Replace this with your preferred output paths
data_path = pathlib.Path("../raster-ml/original_size/coco/images")
json_path = pathlib.Path("../raster-ml/original_size/coco/json")

# Dimensions of the moving window and output images
width, height = 256, 256

# Starting a new CocoDataset
description = "My First Dataset"
contributor = "User"

# version and date_created are automatically set
dataset = create_dataset(description=description, contributor=contributor)
# You can also load existing COCO datasets
# dataset = load_dataset(json_path=json_path)

# Loading GIS data with rasterio and geopandas
labels = gpd.read_file('../raster-ml/original_size/labels_bbox/S39W061-180_20-01-19.geojson')
raster_source = rasterio.open('test.tif')

# (Optional) Apply any spatial or attribute queries here
# labels = labels.loc[labels["ids"].isin([1, 2, 3])]
# labels = labels.loc[labels.within(some_polygon)]

# Find and save all Annotation instances
dataset = append_dataset(
    dataset=dataset,
    images_dir=data_path,
    src=raster_source,
    labels=labels,
    window_bounds=[(width, height)],
    id_attribute=None,  # column with category_id values
    name_attribute="ids",  # column with category_name values
    super_attribute=None,  # optional column with super_category values
)

# Encode CocoDataset instance as JSON and save to json_path
save_dataset(dataset=dataset, json_path=json_path)

Also, is there a way to feed in a pre-chipped dataset (256x256) in order to create the COCO Dataset?

I already have all tif files and corresponding geojson pre-chipped....

[jaspersiebring]

Thanks so much! The variable dimensions should not be a problem but I do wonder about the driver that is being assigned.

If you don't specify a driver manually, Rasterio will just assign one based on the file's extension which appears to be JPEG. This driver is then also used throughout other operations like the clipping which assumes a georeferenced file/appropriate driver (which I suspect is not the case right now). Can you verify the file extension?

JPEG is supported by GeoCOCO through GDAL but they're not all created equally (see drivers). You can try manually specifying a different driver when you run rasterio.open(driver=some_driver) with some_driver being anything from rasterio.drivers().

How were these images created? I can try it for myself if you are okay with sending me a sample (won't do anything else with it).

[mthsdiniz-usp]

Hey Jasper, i'm feeding a tif file to rasterio in order to produce the COCO dataset, but even so it's trying to output a JPEG. If it can output a tif file without even creating patches that would be even better.

Here are a sample of the files i'm using:
image.zip

image.tif = 18 band imagery 256 x 256
label.geojson = rectangular polygon around my objects
label.tif = binary mask for each one of the objects

ps: i zipped the files because github does not accept the formats.

Thanks a lot for the help!