nyan

nyan (not your average name) is an image pre-processing and post-processing library built on top of Pillow, numpy OpenCV-Python, and imageio.

The library is designed for Computer Vision tasks, where you (i) pre-process, (ii) perform inference producing bounding boxes or segmentation masks, and (iii) transform those results back to the original image space for visualisation with the image.

It is also useful for jointly applying pre-processing to polygons, bounding boxes or segmentation masks and the original image or video.

Note: this library is still under development and is incomplete, feel free to experiment with it in it's current state, but there is a long list of TODOs before the first stable release.

TODO

• Implement minimal features for (i) pre-processing, (ii) colour conversions, and (ii) mapping visualisation objects.
• Implement debug_mode
• Add minimal README
• Add minimal example for segmentation
• Add minimal visualisation utilities
• Complete tests
• Add documentation
• Test and/or add support on/for windows and mac
• First stable release on PyPi
• Add TFRecords support
• Add Keras generator support
• ...

Installation

To install, simply clone this repository and use pip to install.

git clone https://github.com/TWRogers/nyan.git
pip install -e nyan

Example for segmentation

To run the segmentation example, run example_masks.py:

python ./examples/example_masks.py

The example takes an example video and performs arbitrary transformations to it, visualising the intermediate steps using the debug_mode feature.

Next, it loads a collection of segmentation masks ...

... and transforms them into the same space as the pre-processed video

Finally, it transforms the mask back to the original space as a sanity check