Computer Vision algorithms are frequently formulated as linear algebra problems in a high dimensional space, where each asset is stripped into a vector. In this high dimensional space we may perform any number of operations, but normally we can’t stay in this space for the whole algorithm - we normally have to recast the vector back into it’s original domain in order to perform other operations.
An example of this might be seen with images, where the gradient of the intensity values of an image needs to be taken. This is a complex problem to solve in a vector space representation of the image, but trivial to solve in the image domain.
Menpo bridges the gap by naively supporting bi-directional vectorisation of
it’s types through the
Vectorizable interface. Through this, any type can
be safely and efficiently converted to a vector form and back again. You’ll find
the key methods of
Vectorizable are extensively used in Menpo. They are
as_vector- generate a vector from one of our types.
from_vector- rebuild one of our types from a vector
from_vector_inplace- alter an object inplace to take on the new state
1. Each type defines it’s own form of vectorization. Calling
as_vector on a
Image returns all of the pixels in a single strip,
whilst on a
MaskedImage only the true pixels are returned. This
distinction means that much of Menpo’s image algorithms work equally well with
masked or unmasked data - it’s the
Vectorizable interface that abstracts
away the difference between the two.
2. Lots of things are vectorizable, not just images. Pointclouds and lots of transforms are too.
3. The length of the resulting vector of a type can be found by querying the ``n_parameters`` property.
4. The vectorized form of an object does not have to be ‘complete’.
from_vector_inplace can use the object they are
called on to rebuild a complete state. Think of vectorization more as a
parametrization of the object, not a complete serialization.