Sharing features for multi-class object detection

Antonio Torralba, Kevin Murphy and Bill Freeman

MITComputer Science

and Artificial Intelligence Laboratory (CSAIL)Sept. 1, 2004

The lead author: Antonio Torralba

Goal

• We want a machine to be able to identify thousands of different objects as it looks around the world.

Multi-class object detection:

Localfeatures no car

ClassifierP( car | vp )

Vp

no cowClassifier

P( cow | vp )

no personClassifier

P(person | vp )

…

Bookshelf

Desk

Screen

Desired detector outputs:

One patch

Need to detect as well as to recognize

Detection: localize in the image screens, keyboards and bottles

In detection, the big problem is to differentiate the sparse objects against the background

Recognition: name each object

In recognition, the problem is to discriminate between objects.

There are efficient solutions for

detecting a single object category and view:

Viola & Jones 2001; Papageorgiou & Poggio 2000, …

Lowe, 1999

detecting particular objects:.

But the problem of multi-class and multi-view object detection is still largely unsolved.

Leibe & Schiele, 2003;

detecting objects in isolation

Why multi-object detection is a hard problem

viewpoints

Need to detect Nclasses * Nviews * Nstyles.Lots of variability within classes, and across viewpoints.

Object classes

Styles, lighting conditions, etc, etc, etc…

Existing approaches for multiclass object detection (vision community)

Using a set of independent binary classifiers is the dominant strategy:

• Viola-Jones extension for dealing with rotations

- two cascades for each view

• Schneiderman-Kanade multiclass object detection

a) One detector foreach class

Promising approaches…• Fei-Fei, Fergus, & Perona, 2003

Look for a vocabularyof edges that reducesthe number of features

• Krempp, Geman, & Amit, 2002

Characteristics of one-vs-all multiclass approaches: cost

Computational cost grows linearly with Nclasses * Nviews * Nstyles …Surely, this will not scale well to 30,000 object classes.

Characteristics of one-vs-all multiclass approaches: representation

Some of these parts can only be used for this object.

What is the best representation to detect a traffic sign?

Very regular object: template matching will do the job

Parts derived fromtraining a binary classifier.

~100% detection ratewith 0 false alarms

Meaningful parts, in one-vs-all approaches

Part-based object representation (looking for meaningful parts):

• A. Agarwal and D. Roth

Ullman, Vidal-Naquet, and Sali, 2004: features of intermediate complexity are most informative for (single-object) classification.

• M. Weber, M. Welling and P. Perona

…

Multi-class classifiers (machine learning community)

•Error correcting output codes (Dietterich & Bakiri, 1995; …)

But only use classification decisions (1/-1), not real values.

•Reducing multi-class to binary (Allwein et al, 2000)Showed that the best code matrix is problem-dependent; don’t address how to design code matrix.

• Bunching algorithm (Dekel and Singer, 2002)Also learns code matrix and classifies, but more complicated than our algorithm and not applied to object detection.

• Multitask learning (Caruana, 1997; …)

Train tasks in parallel to improve generalization, share features. But not applied to object detection, nor in a boosting framework.

Our approach

• Share features across objects, automatically selecting the best sharing pattern.

• Benefits of shared features: – Efficiency– Accuracy– Generalization ability

Algorithm goals, for object recognition.

We want to find the vocabulary of parts that can be shared

We want share across different objects generic knowledge about detecting objects (eg, from the background).

We want to share computations across classes so that computational cost < O(Number of classes)

Object class 1

Object class 2

Object class 3

Object class 4

Total number of hyperplanes (features): 4 x 6 = 24. Scales linearly with number of classes

Independentfeatures

Total number of shared hyperplanes (features): 8

May scale sub-linearly with number of classes, and may generalize better.

Sharedfeatures

Note: sharing is a graph, not a tree

bR

3b

3

3

Objects: {R, b, 3}

This defines a vocabulary of parts shared across objects

At the algorithmic level

• Our approach is a variation on boosting that allows for sharing features in a natural way.

• So let’s review boosting (ada-boost demo)

Boosting demo

Joint boosting, outside of the context of images.

Additive models for classification

+1/-1 classificationclasses

feature responses

Feature sharing in additive models

1) Simple to have sharing between additive models

2) Each term hm can be mapped to a single feature

H1 = G1,2 + G1

H2 = G1,2 + G2

Flavors of boosting

• Different boosting algorithms use different lossfunctions or minimization procedures (Freund & Shapire, 1995; Friedman, Hastie, Tibshhirani, 1998).

• We base our approach on Gentle boosting: learns faster than others(Friedman, Hastie, Tibshhirani, 1998; Lienahart, Kuranov, & Pisarevsky, 2003).

Joint Boosting

We use the exponential multi-class cost function

At each boosting round, we add a function:

classes

classifier output for class c

membership in class c,+1/-1

Newton’s method

Treat hm as a perturbation, and expand loss J to second order in hm

classifier with perturbation

squared error

reweighting

Joint Boosting

weight squared errorWeight squared error over training data

a+b

b vf

Given a sharing pattern, the decision stump parameters

are obtained analytically

Feature output, v

For a trial sharing pattern, set weak learner parameters to optimize overall classification

hm (v,c)

Response histograms for background (blue) and class members (red)

hm(v,c)

kc=1

kc=2

kc=5

The constants k prevent sharing

features just due to an asymmetry between

positive and negative examples for each

class. They only appear during

training.

Joint Boosting: select sharing pattern and weak learner to minimize cost.

Algorithm details in CVPR 2004, Torralba, Murphy & Freeman

Approximate best sharingBut this requires exploring 2C –1 possible sharing patterns

Instead we use a first-best search:

• S = []

• 1) We fit stumps for each class independently

• 2) take best class - ci

• S = [S ci]

• fit stumps for [S ci] with ci not in S

• go to 2, until length(S) = Nclasses

• 3) select the sharing with smallest WLS error

Effect of pattern of feature sharing on number of features required (synthetic example)

Effect of pattern of feature sharing on number of features required (synthetic example)

2-d synthetic example

3 classes + 1 background class

No feature sharing

Three one-vs-all binary classifiers

This is with only 8 separation lines

With feature sharing

This is with only 8 separation lines

Some lines can be shared across classes.

The shared features

Comparison of the classifiersShared features: note better isolation of individual classes.

Non-shared features.

Now, apply this to images.

Image features (weak learners)

32x32 training image of an object

12x12 patch

Location of that patch within the 32x32 object

gf(x) Mean = 0 Energy = 1wf(x) Binary mask

Feature output

The candidate featurestemplate position

The candidate features

Dictionary of 2000 candidate patches and position masks, randomly sampled from the training images

template position

Database of 2500 images

Annotated instances

Multiclass object detection

21 objects

We use [20, 50] training samples per object, and about 20 times as many background examples as object examples.

Feature sharing at each boosting round during training

Feature sharing at each boosting round during training

Example shared feature (weak classifier)

At each round of running joint boosting on training set we geta feature and a sharing pattern.

Response histograms for

background (blue) and class members (red)

Non-shared feature Shared feature

Non-shared feature Shared feature

How the features were shared across objects (features sorted left-to-right from generic to specific)

Performance evaluation

Area under ROC (shown is .9)

False alarm rate

Cor

rect

det

ecti

on r

ate

Performance improvement over training

Significant benefit to sharing features using joint boosting.

ROC curves for our 21 object database

• How will this work under training-starved or feature-starved conditions?

• Presumably, in the real world, we will always be starved for training data and for features.

70 features, 20 training examples (left)

Shared features Non-shared features

15 features, 20 training examples (mid)70 features, 20 training examples (left)

Shared features Non-shared features

15 features, 2 training examples (right)15 features, 20 training examples (middle)

70 features, 20 training examples (left)

Shared features Non-shared features

Scaling

Joint Boosting shows sub-linear scaling of features with objects (for area under ROC = 0.9).

Results averaged over 8 training sets,

and different combinations of

objects. Error bars show variability.

Red: shared features Blue: independent features

Red: shared features Blue: independent features

Examples of correct detections

What makes good features?

• Depends on whether we are doing single-class or multi-class detection…

Generic vs. specific features

Parts derived fromtraining a binary classifier.

In both cases ~100% detection rate with 0 false alarms

Parts derived from training a joint classifier with 20 more objects.

Qualitative comparison of features, for single-class and multi-class detectors

Multi-view object detectiontrain for object and orientation

Viewinvariantfeatures

Viewspecificfeatures

Sharing features is a natural approach to view-invariant object detection.

Multi-view object detection

Sharing is not a tree. Depends also on 3D symmetries.

…

…

Multi-view object detection

Multi-view object detection

Strong learner H response for car as function of assumed view angle

Visual summary…

Features

Object units

Summary

• Argued that feature sharing will be an essential part of scaling up object detection to hundreds or thousands of objects (and viewpoints).

• We introduced joint boosting, a generalization to boosting that incorporates feature sharing in a natural way.

• Initial results (up to 30 objects) show the desired scaling behavior for # features vs # objects.

• The shared features are observed to generalize better, allowing learning from fewer examples, using fewer features.

end