Constructing Category Hierarchies for Visual Recognition

1
Constructing Category Hierarchies for Visual
Recognition

• Marcin Marszaklek and Cordelia Schmid

2
Introduction

• Hierarchical classification scales well in the
  number of classes
• O(n2) one-vs-one
• O(n) one-vs-rest
• O(log(n) classification tree
• Previous works to construct class hierarchies
• By hand Zweig07
• From external sources Marszaklek07
• From visual similarities
• Exhaustive Yuan06
• Top-down Chen04, Griffin08
• Botton-up Zhigang05, Griffin08

3
Motivation Disjoint VS overlap

• Previous works disjoint partitioning of classes
  (class separability)
• Increasingly difficult to disjoint partition for
  large number of classes.

• Propose relaxed hierarchy postpone uncertain
  classification decisions until the number of
  classes get reduced and learning good decision
  boundaries becomes tractable.

4
Method

• Building relaxed hierarchy
• Train top-down classifiers using hierarchy

5
Building top-down relaxed hierarchy

• Using balanced Normalized-cut, split the set of
  classes such that
• Further relaxation
• Find the class on boundary
• Define the split (a overlap ratio)

, given a partition
6
Building top-down relaxed hierarchy conti.
7
Train/test top-down classifiers

• Training hierarchy
• For each node of DAG, samples of Ln\Rn as
  positive sample, Rn\Ln as negative samples
• Samples in classes XnLn ? Rn not for training
• Testing
• Traversal DAG until a leaf is reached.
• The decision is either directly the class label
  (leaves containing one class), or performing OAR
  classification on the remaining classes in
  current leaf.

8
Results one-vs-rest
Confusion between mountain/touring bikes
High intra-class variability
Low intra-class variability
9
Class hierarchies caltech 256
hand-crafted hierarchy
relaxed hierarchy
Disjoint visual hierarchy
Categories animal, natural phenomena and
man-made objects
10
Results
Average per-class accuracy on Caltech-256
11
Results conti.
Complexity in the number of classes r relaxed
training sample per class.
Speed-for-accuracy trade-off
12
Learning and Using Taxonomies For Fast Visual
Categorization

• Gregory Griffin and Pietro Perona

13
Motivation
expensive
Given a testing sample
O( category)
inexpensive
O( log2( category))
One-vs-rest strategy VS hierarchical strategy
14
Methods

• Building confusion matrix
• Building Taxonomies
• Re-train top-down classifiers

15
Building confusion matrix

• Multi-class classification one-vs-rest strategy
• Classifier Spatial Pyramid Matching
• Training data only and loo validation

16
Building Taxonomies

• Intuition
• Categories that are easily confused should be
  grouped together
• Decisions between easily-confused categories
  sholuld be taken later in the decision tree.
• Method
• Self-Tuning Spectral Clustering
• Greedy, bottom-up grouping using mutual infor.

17
Re-train top-down classifiers

• Known the taxonomy tree of categories as a binary
  tree
• At each node, reformulating a binary-classifier
• Again, using Spatial Pyramid Matching SVM
• F_train 10

18
Results
Red insects Yellow birds Green land
mammals Blue aquatic mammals
Taxonomy tree for Caltech-256
19
Trade-off between performance and speed
Spectral clustering
Greedy clustering
A ordinary one-vs-rest multi-classifier C each
testing image goes through the tree B
intermediate level
N_train 10 5x speed up with 10 performance
drop
20
Results
Cascade performance / speed trade off as a
function of training example/class 20x speed up
with 10 performance drop for N_train50