NNNN

Nerf or Nothin' Neural Network

The universal binary image classifier.

Setup

With conda (4.8.2):

conda env create -f env.yml
conda activate NNNN

Making Ground Truth

A workflow is available to create new labeled data. Save images to data/. The script util/labeler.py uses OpenCV to allow the user to select bounding rectangle around any / all / no targets in the images. Images are then downsampled such that max(height, width) = 500 and updated within data/. Bounding rectangle information is saved to data/truth/ground_truth_0.csv, and original images are saved to data/originals/.

A small sample of true negative images are included in data/, taken from the PASCAL Visual Object Classes Challenge.

To label new data, simply save the images to data/ and run:

cd util
python labeler.py

This will only offer images not in data/truth/ground_truth_0.csv for bounding rectangle selection.

To re-label data or label a specific new image (or both), pass the -f argument with the file names:

cd util
python labeler.py -f 0.jpeg 3.jpeg

TODO:

• Do the actual work of building / training model.