1
C H A P T E R 1
Introduction
Designing visual features has been a fundamental problem of computer vision for many decades.
Visual features have to represent visual information in a suitable way, where the definition of
suitability has been shifting regularly, resulting in various feature design principles. Also, after
the recent progress of deep learning, and deep features, these principles are still relevant for
understanding and improving deep learning functionality and methodology. A principled un-
derstanding of feature design is the basis for analyzing the lower layers in deep networks, trained
on complex tasks and big datasets.
1.1 FEATURE DESIGN
Feature design principles are systematic approaches to design features and avoid ad-hoc solu-
tions (“pixel f***ing” according to Koenderink and van Doorn [2002]). e idea behind visual
features is to transform raw values from cameras and other visual sensors into some interme-
diate representation, the feature descriptor, that is more relevant to the solution of the problem
addressed [Koenderink, 1993].
Historically, the design of feature descriptors has seen many changes. During the early
years of computer vision, feature extraction has mainly been considered as an algorithmic prob-
lem, a sub-problem of artificial intelligence, within computer science [Papert, 1966]. Since then,
computer vision has become an interdisciplinary field and feature extraction has advanced to-
ward models influenced by physics, statistics, electrical engineering, mathematics, neuroscience,
and cognitive science.
In this process, the interpretation of an image has changed from being a simple byte
array to being a sophisticated object reflecting real-world properties including models for point-
spread functions and lens distortion, noise and outliers, continuous signals with sampling and
interpolation, perspective transformations, and reflectance modeling. Even the interpretation by
observers is sometimes reflected in feature extraction, e.g., by non-maximum suppression and
local inhibition, adaptivity, and structural completion. Still, all features are based on digital image
information and need to be computed in a digital system taking into account computational and
memory resources.
Visual features are usually considered as part of some hierarchical processing scheme. Already
in the work of Marr [1982], visual information undergoes processing in several steps, starting
at the primal sketch (basically feature extraction) and ending at a 3D model representation.
Later, hierarchical models split the feature extraction into several levels, low-level features such
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.