Dataset for training, optimising and evaluating segmentation algorithms

[quicklizard99]

We need to improve the segmentation algorithm and have more formal evaluation with a great variety of object types. This requires knowledge of correct bounding boxes.

As a first pass, we should segment and refine some or all of the 804 JPEGs at https://dx.doi.org/10.5519/0018537, and augment these data with images of other object types.