\section{Notes}

Restricting to using only the top scoring features within a query annotation seems to work well. 

It has to do with this idea of distinction. 

There are features that might not even show up. 

What is the probability that a feature even exists in the database?


Zebras have two types of features (that lie on a spectrum)

(1) Those that distinguish them from other species, and
(2) Those that distinguish themselves from other zebras.

This is an oversimplification

A fine-grained class is a subclass of a superclass.
We can view the superclass (zebra in species) as the parent in a
hierarchical taxonomy.
This parent is a class in the parent-class-space.
The goal of fine-grained recognition is to classify an object as a
subclass --- a child node of the parent class --- (eg fred the zebra).
For us fine-grained class is in the target-class-space.

More generally, for preforming fine-grained recognition, an objects features
can be partitioned into two types of features:

(1) Features that distinguish it in the parent-class space.
(2) Features that distinguish it in the target-class space.

All types of classification call into this structure.
In fine grained species recognition the parent-class space might be
animal-species, while the target-class space might be birds.
Given that this is a bird species (of the parent-class space), classify
the subspecies.

For individual identification the parent-class species is the sub-species
(eg Grevy's zebra / plains zebra).
The target class is the individual.

Only the features that distinguish between classes in the target-class
space are useful in any sort of classification.
If feature extraction includes features from the parent-class space it
will cause confusion in the classifier unless the classifier is built in
such a way that it can ignore such features.


However, features in the parent-class space are not exactly mutually
exclusive of features in the target-class space.
A stripe might be indicative of a zebra, but depending on the subtleties
of the shape it may individually identify it as well.


This argues that we need feature extraction that can accurately describe
such subtleties.


\paragraph{What do do when features aren't there?}

At some point --- enough --- distinctive descriptors have been matched, so it doesn't bother the classifier that 
you haven't matched more.

This argues again for taking the top X most distinctive descriptors.


\paragraph{Nearest neighbors as bag-of-words}

If we treat every descriptor in the database as its own word in a bag of words,
we can compute the bag of words vector for a query by assigning each query
descriptor to its k nearest neighbors.


Idea for chapter 5:

System Structure

Dependency Cache

How I operate the system.

Things that I did with the system:

\paragraph{Tags}

I tagged several cases as ``interesting''. 

Talk about other tags used.


\paragraph{Advantages of Naive Bayes}
% https://www.quora.com/Classification-machine-learning/When-should-I-use-a-K-NN-classifier-over-a-Naive-Bayes-classifier
Handles missing data in the query image.  Although, it does not handle missing data from the database image.



\paragraph{Experiment Ideas}
How much worse do results get when you remove the best matching annotation?
How fast?

\paragraph{VsOne Ideas}

Spatial verification produces false correspondences by matching a large random distribution of keypoints 
to a small random distribution of keypoints. 

We need a way to discourage scale change in the estimated transformation because the
scale of both annotations should be more or less the same? 

This obviously is not always true.
But we can use a configuration to get the majority of the cases when it is
  true.

\paragraph{Viewpoint co-visiblity ideas}
    With the animals of a particular species we can define a common set of
      viewpoints and poses that all members of that species can be seen in.
    \begin{itemize}
        \item a high degree of distinctive co-visible distinctive features implies identity-equivalence with high probability
        \item a low degree of distinctive co-visible distinctive features does not imply identity-inequivalence 
        \item a low degree of co-visible features from the same viewpoint implies identity-inequivalence with high probability
        \item identity-equivalence is transitive
        \item non-compatible (disjoint) viewpoints implies no co-visibility
    \end{itemize}
    % define:
    % identity-equivilence
    % identity-inequivilence
    % identity-indeterminacy
    % co-visible
    %This classification puts the system in a state of error until 
    %Inversely, if two unoccluded regions of the same viewpoints do not have
    %  any visual similarity then name-inequivalence can be concluded.


\paragraph{Determining False Negatives}
Look at names that only have one encounter.
If this number grows as the database size grows, it is evidence that the
  algorithm is missing matches.
This might be the case for some plains zebras which are not distinctive.


\paragraph{Space}

Once you are so far from any point in space, you can only be so certain about where you are. 
You need to be at least close to some point.

once you are so far from any point in space, you can only be so certain about where you are. 


\paragraph{
Naive Bayes works so well because it is a voting procedure.
The idea of voting is intuitive, natural, and easily understood.
If all the members of the population need to come to a consensus on something, 
voting is the only procedure (apart from violence, intuitively speaking)
that can solve this problem.
But who gets to vote and what each agents preferences and prerogative are is not always clear.


}