/cocofy

Primary LanguagePython

Helper functions to convert bounding box data organized in dataframes or other formats to COCO format annotations.

Assuming you have data in similar format to the files found here, and assuming you have a column in your dataframe named filename, you can generate COCO annotations in the following way:

df = pd.read_parquet("design5.parquet")

df["iscrowd"] = 0
df["ignore"] = 0

# df["filename"] # Assumed to exist
# df["image_width"] = 1000 # Assumed to exist
# df["image_height"] = 1600 # Assumed to exist

df = df.rename(columns={"label_xgboost_adjusted": "label"})

# Output annotations for the 3 labels nyheter_sverige, ledare and debatt
cocofy_from_df(df_test, 
               categories=["nyheter_sverige", "ledare", "debatt"], 
               image_folder="images")

Documentation

A COCO conversion tool creating a minimally viable annotations.json for object detection from a dataframe. Also copies image files from source image folder to destination folder.

  • df (pd.DataFrame): Dataframe with the following mandatory columns:

    • filename (base name of file without folders)
    • x
    • y
    • width
    • height
    • image_width (image file width)
    • image_height (image file_height
    • label
    • iscrowd (see COCO docs, set to 0 if unsure)
    • ignore (see COCO docs, set to 0 if unsure)

  • categories (list): List with category names to be included as annotations. Categories which are not listed will be ignored when creating bbox annotations. Category names need to match label names in dataframe.

  • image_folder (str): Source folder where your images are located. The filenames present in df will be copied over to a folder named "cocofy/images" in your current working directory.

  • destination_folder (str): Folder where your annotations and images will be stored. Default: "cocofy". I.e. annotations.json will be placed in folder "cocofy", and images will be under "cocofy/images".

  • info (None | dict): Optional dictionary with information about your dataset. E.g. {"year": 2021, "version": "1.0", "contributor": "John Smith"}.

  • copy_images (bool): If True, copies images from image_folder to {destination_folder}/images