thumbor_extras

Some useful extensions to thumbor - extra filters and detectors.

pip install thumbor_extras

Detectors

Add these detectors to your thumbor.conf:

DETECTORS = [
    'thumbor_extras.detectors.dnn_face_detector',
    'thumbor_extras.detectors.dnn_object_detector'
]

dnn_face_detector

http://localhost:8888/unsafe/smart/filters:draw_center_of_mass():draw_focal_points(2,.25)/https://farm2.static.flickr.com/1098/998604755_c2155161b3.jpg

An improved face detector that uses a deep neural network. More specifically, it uses the SSD object detection model, trained specifically on faces. According to the OpenCV repo, it was trained with "some huge and available online dataset." :/

For more information on how the normal thumbor face detector compares to this one in both accuracy and efficiency, see this comparison script.

dnn_object_detector

http://localhost:8888/unsafe/smart/filters:draw_center_of_mass():draw_focal_points(2)/https://farm2.static.flickr.com/1308/562554257_53f458e042.jpg

The full SSD object detection model, trained on the COCO dataset. It was built as a MobileNet, a class of efficient, light weight deep neural network models meant to work well in mobile and embedded applications.

Filters

Add these filters to your available filters in thumbor.conf:

FILTERS = [
    ...other filters,
    'thumbor_extras.filters.draw_center_of_mass',
    'thumbor_extras.filters.draw_focal_points'
]

draw_center_of_mass

Draws a circle at the calculated center of mass, according to the focal points.

Usage: draw_center_of_mass([r, g, b])

Examples:

• draw_center_of_mass(): draws a circle with default red color.
• draw_center_of_mass(0,0,255): draws a blue circle.

draw_focal_points

Draws a box around the focal points, for displaying the results of the detectors on a given image. Takes optional arguments for box color and line width.

Usage: draw_focal_points([line_width, label_height_percentage, show_heatmap, show_labels, show_rainbow, r, g, b])

Arguments:

• line_width: the width of the box lines
• label_height_percentage: the height of the class label as a percentage of the box
• show_heatmap: make the box more opaque for higher-confidence detections, more transparent for less confident detections
• show_labels: print the class label at the top of the box
• show_rainbow: use a different color for each focal point box; pass false to provide a single color RGB value for all boxes
• r: R component of RGB color of box, default 0
• g: G component of RGB color of box, default 255
• b: B component of RGB color of box, default 0

Examples:

• draw_focal_points(): draws boxes with defaults (each box a different color, with opacity according to the detection confidence).
• draw_focal_points(5,0,false,false,false,255,0,0): draws solid red boxes with 5 pixel line width, and no class labels.

TODO

• Choose thresholds empirically

Development

Because I will definitely forget. Python development can be a pain.

# -- ENV --
# use python 2 - thumbor doesn't support 3 yet.
virtualenv thumbor_env --python=/path/to/python2
source thumbor_env/bin/activate

# -- DEVELOP --
pip install -e .

# -- TEST --
# i don't know why setup.py can't install pytest...but it fails without manual install.
# it could be a python2 / mac thing.
pip install pytest
python setup.py test

# -- DISTRIBUTE --
python setup.py sdist
# make a wheel from the source tarball (to use manifest correctly)
pip wheel --no-index --no-deps --wheel-dir dist dist/*.tar.gz
pip install twine
# make sure pip URLs are configured per https://truveris.github.io/articles/configuring-pypirc/
python -m twine upload --repository pypi dist/*