Hieralb: Hierarchical Album

A niche package

Hieralb is composed of a small set of Python functions that implement a consistent identification scheme for images used in visual categorization. In this package, a collection of images is referred to as an album. Within an album, a given image can often be categorized at multiple levels (e.g., bird, kingfisher, ringed kingfisher) creating a hierarchy of categories. In many applications, it is desirable to consider these different levels of categorization. Hieralb enables you to easily generate unique image IDs and class IDs for all possible categories. Hieralb generates a set of plain text files that specify the relationship between the actual images, image IDs and class IDs. Lastly, Hieralb provides a Python class so that the album's structural relationships and ID's can easily be used in other Python applications.

Getting started

These instructions provide a guick overview of what you need to do to get started. More detailed instructions can be found below.

1. Create a new folder <my_album_folder>.
2. Create the folder images inside <my_album_folder>.
3. Inside images, create appropriately named folders. The folder names will become the literal class names.
4. Place all of your images in the appropriate folders.
5. To create the appropriate album files run:$ python create_album_files.py <my_album_folder> <my_album_prefix>

Installation

There are two ways to get Hieralb:

1. Install using pip: pip install hieralb
2. Clone from Git Hub: https://github.com/roads/hieralb.git

Detailed instructions

This file covers the creation and management of domain directories for use in applications that require organized image databases. The core organization principle is that an image can be classified at different levels. Classification levels are modeled using a hierarchical folder structure.

For example, consider an album containing bird images. Let's call this album birds-9, since there are nine different species. Let us also assume that we want to capture two different levels of categorization: the taxonomic family and taxonomic species. The bird directory might be organized in the following way:

birds-9/
└── bird/
    ├── Parulidae/
    │   ├── Bobolink/
    │   ├── Hooded_Oriole/
    │   └── Scott_Oriole/
    ├── Passeridae/
    │   ├── Hooded_Warbler/
    │   ├── Kentucky_Warbler/
    │   └── Magnolia Warbler/
    └── Icteridae/
        ├── Chipping_Sparrow/
        ├── Fox_Sparrow/
        └── Harris_Sparrow/

After arranging the image/ directory to your satisfaction (with the actual images placed in the appropriate species folder), you can run: $ python create_album_files.py birds-9 10000000

After running the script, the album directory will contain (at least) the following three text files:

• classes.txt

  • A file listing all the classes (from all hierarchical levels) along with their corresponding class_id.
  • The file is composed of two columns using a space delimeter: <class_id> <class_name>
  • The class_id begins at 0 for each album, thus the class ID is not unique across albums. The root class in the directory tree is assigned class_id 0.

• images.txt

  • A file listing all the images and their corresponding imag_id.
  • The file is composed of two columns using a space delimeter: <unqiue_image_id> <image_filename>
  • The unique_image_id is constructed using a numerical prefix supplied by the user (e.g., 10000000) which is added to a value starting at 1 (e.g., 10000001, 10000002, 10000002, ...). By using a prefix, the unique_image_id is unique across albums. Use the prefix "0" if you don't want to use a prefix.

• image_class0.txt

  • A file mapping each image to its finest-grained class.
  • The file is composed of two columns using a space delimeter: <unique_image_id> <class_id>

• class_class.txt

  • A file listing the mapping between a child class and its parent class.
  • A file composed of two columns using a space delimeter: <child class_id> <parent class_id>

If the image/ directory is more than one level deep (i.e, it contains nested classes), then additional files will be created. These files will follow the structure:

• class<n>_class<n+1>.txt

  • A file listing the mapping between a class level n and its parent level n+1.
  • A file composed of two columns using a space delimeter: <child class_id> <parent class_id>
  • Level zero starts at the deepest leaf node of the directory tree.

Once the appropriate files have been created, you can use the Album class to load the data associated with an album into convenient python object. An album object can be instantied by importing passing a string containing the path to the album directory.

from hieralb.core import Album
birds_9_album = Album('\birds-9')

Conventions and assumptions

• Code has only been tested using Python 3
• Folder names should be singular within the “images” folder (e.g., use bird not birds).
• Use underscore as a placeholder for whitespace (e.g., “Hooded_Warbler”, not

“Hooded Warbler”). Actual spaces can not be used in folder or filenames. - The depth of each leaf node (from the parent node) must be the same.

Versioning

This package folows the Semantic Versioning 2.0.0 rules (https://semver.org).

Authors

• Brett D. Roads

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE.txt file for details.