Monty.
Copy Markdown
View Source
Multi-scale images, and the comparisons VOCUS2 makes across them.
Mirrors tbp/monty/frameworks/models/salience/strategies/vocus2/pyramids.py.
An attention model has to find what stands out at some size and not at
others: a small bright patch, a long edge, a dark corner. It does that by
building the image at several resolutions (octaves), each blurred several times
(scales), and then comparing neighbouring scales — a centre against a
surround — which is what center_surround/2 and laplacian/1 are for.
The pyramid follows Lowe (2004) as VOCUS2 implements it: each octave starts
from the extra scale the previous one computed, resampled by :area, and each
following scale is blurred by whatever it takes to reach the next sigma from
the one before — sqrt(next² - previous²), because blurs compose in
quadrature. That is the step that makes a pyramid a pyramid rather than a
handful of blurs.
Summary
Functions
A centre pyramid and its surround, at the same scales.
Collapses a pyramid into one image, the shape of its finest scale.
Collapses a pyramid by taking the strongest plane at each pixel.
Collapses a pyramid by averaging its planes at each pixel.
Combines pyramids of the same shape into one, plane by plane with reduce.
Combines pyramids by taking the strongest plane at each pixel.
Combines pyramids by averaging their planes at each pixel.
A Gaussian pyramid: :n_scales blurred versions of the image at each octave.
The Laplacian pyramid: each octave minus the octave below it, enlarged back.
The resolutions an image of image_shape is built at, finest first.
Types
@type plane() :: [[float()]]
Functions
@spec center_surround(plane(), keyword()) :: {Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t(), Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()}
A centre pyramid and its surround, at the same scales.
The surround is the centre blurred further, by the sigma that takes a centre
of :center_sigma to a surround of :surround_sigma — in quadrature again,
and scaled up with each scale so a coarse level is not compared against a fine
one's surroundings.
@spec collapse( Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t(), ([plane()] -> plane()) ) :: plane()
Collapses a pyramid into one image, the shape of its finest scale.
Every plane is enlarged to that shape by :cubic and then reduced with
reduce — so a map that exists at several scales becomes one map again.
@spec collapse_max( Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t() ) :: plane()
Collapses a pyramid by taking the strongest plane at each pixel.
@spec collapse_mean( Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t() ) :: plane()
Collapses a pyramid by averaging its planes at each pixel.
@spec combine( [Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()], ([plane()] -> plane()) ) :: Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()
Combines pyramids of the same shape into one, plane by plane with reduce.
A single pyramid is already the answer, which is how a config that names one map and not several still works.
@spec combine_max([ Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t() ]) :: Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()
Combines pyramids by taking the strongest plane at each pixel.
@spec combine_mean([ Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t() ]) :: Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()
Combines pyramids by averaging their planes at each pixel.
@spec gaussian_pyramid(plane(), keyword()) :: Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()
A Gaussian pyramid: :n_scales blurred versions of the image at each octave.
:sigma is the blur of the first scale of the first octave, and every octave
starts from the same sigma, so the scales of one octave and the next mean the
same thing.
@spec laplacian( Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t() ) :: Monty.Frameworks.Models.Salience.Strategies.Vocus2.Pyramids.Pyramid.t()
The Laplacian pyramid: each octave minus the octave below it, enlarged back.
What survives the subtraction is what the coarser octave could not represent — the detail this resolution adds — which is what makes an edge visible at the scale it belongs to. Has one fewer octave than its input.
@spec octave_shapes({pos_integer(), pos_integer()}, pos_integer() | nil) :: [ {pos_integer(), pos_integer()} ]
The resolutions an image of image_shape is built at, finest first.
Each octave is half the previous, and there are as many as fit until a
dimension would fall below one pixel — or fewer, if :max_octaves says so.