This repository holds data/scripts related to the Micro pasts project.

jupyter_notebooks holds IPython/Jupyter notebooks.

scripts hosts useful scripts.

The Python scripts require some third-party libraries. These are listed in the requirements.txt file and can be installed via:

$ pip install -r requirements.txt

In order to train and test a classifier:

First call "%run stackImages path to image folder" to concatenate the image and mask set for classification. E.g: "%run stackImages C:\Python34\palstaves2\2013T482_Lower_Hardres_Canterbury\Axe1"

Next call "%run createClassifier path to image folder classifier type" to train the classifier on the ".npz" file created by "stackImages.py". The same path should be used as for "stackImages" as this is where the ".npz" file is saved. The type of classifier must also be specified, currently only LinearSVC is available.

Finally call "%run useClassifier path to image folder classifier type" to test the classifier on the test set and produce an average pixel error. Currently the error metric is simply absolute difference summed over all pixel classifications. Path should again be the same as "stackImages" as this is where the classifier will have been saved to from the previous step in pickle form.

stackImages has a parameter inside the script which specifies the amount of sampling of an image and it's mask. Currently the sampler chooses an equal number of foreground (1 in mask) and background (0 in mask) pixels. The (N,3) and (N,) arrays are then saved in the path folder as a ".npz".

createClassifier has parameters for choosing amount of training vs testing data (set to 1:4 as of 31/10/15). An error will be thrown if a classifier is not recognised.