Bag-of-features models

for category classification

Cordelia Schmid
 Category recognition
• Image classification: assigning a class label to the image
Car: present
Cow: present
Bike: not present
Horse: not present
…
 Category
Tasks
recognition
• Image classification: assigning a class label to the image
Car: present
Cow: present
Bike: not present
Horse: not present
…
• Object localization: define the location and the category

L
Location
ti
Car Cow
Category
 Difficulties: within object variations

Variability: Camera position, Illumination,Internal parameters

Within-object variations
Difficulties: within class variations
 Image classification
• Given
Positive training images containing an object class

Negative training images that don’t

• Classify
A test image as to whether it contains the object class or not

?
Bag-of-features
Bag of features – Origin: texture recognition

• Texture is characterized by the repetition of basic elements

or textons

Julesz, 1981; Cula & Dana, 2001; Leung & Malik 2001; Mori, Belongie & Malik, 2001;
Schmid 2001; Varma & Zisserman, 2002, 2003; Lazebnik, Schmid & Ponce, 2003
Bag-of-features
Bag of features – Origin: texture recognition

histogram

Universal texton dictionary

Julesz, 1981; Cula & Dana, 2001; Leung & Malik 2001; Mori, Belongie & Malik, 2001;
Schmid 2001; Varma & Zisserman, 2002, 2003; Lazebnik, Schmid & Ponce, 2003
Bag-of-features
Bag of features – Origin: bag
bag-of-words
of words (text)
• Orderless document representation: frequencies of words
from a dictionary
• Classification to determine document categories

Bag-of-words
Co
Commono 2 0 1 3
People 3 0 0 2
Sculpture 0 1 3 0
… … … … …
 Bag-of-features
Bag of features for image classification

SVM

Extract regions Compute Find clusters Compute distance Classification

descriptors and frequencies matrix

[Csurka et al., ECCV Workshop’04], [Nowak,Jurie&Triggs,ECCV’06],

[Zhang,Marszalek,Lazebnik&Schmid,IJCV’07]
 Bag-of-features
Bag of features for image classification

SVM

Extract regions Compute Find clusters Compute distance Classification

descriptors and frequencies matrix
Step 1 Step 2 Step 3
 Step 1: feature extraction
• Scale-invariant
Scale invariant image regions + SIFT (see previous lecture)
– Affine invariant regions give “too” much invariance
– Rotation invariance for many realistic collections “too”
too much
invariance

• Dense descriptors
– Improve results in the context of categories (for most categories)
– Interest
I t t points
i t do
d nott necessarily
il capture
t “all”
“ ll” features
f t

• Color-based
Color based descriptors

• Shape-based
Shape based descriptors
 Dense features

- Multi-scale dense grid: extraction of small overlapping patches at multiple scales

Computation of the SIFT descriptor for each grid cells
-Computation
-Exp.: Horizontal/vertical step size 3 pixel, scaling factor of 1.2 per level
 Bag-of-features
Bag of features for image classification

SVM

Extract regions Compute Find clusters Compute distance Classification

descriptors and frequencies matrix
Step 1 Step 2 Step 3
Step 2: Quantization

Visual vocabulary

Clustering
 Examples
p for visual words

Airplanes

Motorbikes

Faces

Wild Cats

Leaves

People

Bikes
 Step 2: Quantization
• Cluster descriptors
– K-means
– Gaussian mixture model

• Assign
g each visual word to a cluster
– Hard or soft assignment

• Build frequency histogram

 K-means
K means clustering
• Minimizing
g sum of squared
q Euclidean distances
between points xi and their nearest cluster centers

• Algorithm:
– Randomly y initialize K cluster centers
– Iterate until convergence:
• Assign each data point to the nearest center
• R
Recomputet eachh cluster
l t center t as th
the mean off allll points
i t
assigned to it

• Local minimum, solution dependent on initialization

• Initialization important, run several times, select best

 Gaussian mixture model (GMM)
• Mixture of Gaussians: weighted sum of Gaussians

where
ee
 Hard or soft assignment
• K-means
K means  hard assignment
– Assign to the closest cluster center
– Count number of descriptors assigned to a center

• Gaussian mixture model  soft assignment

g
– Estimate distance to all centers
– Sum over number of descriptors

• Represent image by a frequency histogram

 cy
frrequenc Image representation

…..
codewords

• each image is represented by a vector, typically 1000-4000 dimension,

normalization with L1/L2 norm
• fine grained – represent model instances
• coarse grained – represent object categories
 Bag-of-features
Bag of features for image classification

SVM

Extract regions Compute Find clusters Compute distance Classification

descriptors and frequencies matrix
Step 1 Step 2 Step 3
 Step 3: Classification

• Learn a decision rule (classifier) assigning bag-of-

bag of
features representations of images to different classes

Decision Zebra
boundary
Non-zebra
 Training data
Vectors are histograms, one from each training image

positive negative

Train classifier,e.g.SVM
 Linear classifiers
• Find linear function (hyperplane) to separate positive and
negative
i examples l

x i positive : xi  w  b  0
x i negative : xi  w  b  0

Which hyperplane
is best?
 Linear classifiers - margin
x2
(color)

• G
Generalization
li ti iis nott
good in this case:
x1 (roundness)

x2
(color)
• Better if a margin
is introduced: b/|w|

x1 (roundness)
 Nonlinear SVMs
• Datasets that are linearly separable work out great:

0 x

• But what if the dataset is just too hard?

0 x

• We can map it to a higher

higher-dimensional
dimensional space:
x2

0 x
 Nonlinear SVMs
• General idea: the original input space can always be
mapped to some higher-dimensional feature space
where the training set is separable:

Φ: x → φ(x)
 Nonlinear SVMs

• The kernel trick: instead of explicitly computing the lifting

transformation φ(x), define a kernel function K such that
K(xi ,xj ) = φ(xi ) · φ(xj)
j

• This gives a nonlinear decision boundary in the original

feature
eatu e space:
space

  y K ( x , x)  b
i
i i i
 Kernels for bags of features
N
• Histogram intersection kernel: I (h1 , h2 )   min(h (i), h (i))
i 1
1 2

• Generalized Gaussian kernel:

 1 2
K (h1 , h2 )  exp  D(h1 , h2 ) 
 A 
• D can be Euclidean distance  RBF kernel

• D can be χ2 distance
N
D(h1 , h2 )  
h1 (i)  h2 (i) 2
i 1 h1 (i )  h2 (i )
 Combining features
•SVM with multi-channel chi-square kernel

● Channel c is a combination of detector, descriptor

● Dc (Hi , Hj ) is the chi-square distance between histograms

1 m
Dc ( H1 , H 2 ) 
2
i 1
[ ( h1i  h2i ) 2
(h1i  h2i )]

● Ac is the mean value of the distances between all training sample

● Extension: learning of the weights, for example with Multiple

Kernel Learning (MKL)

[J. Zhang, M. Marszalek, S. Lazebnik and C. Schmid. Local features and kernels for
classification of texture and object categories: a comprehensive study, IJCV 2007]
 Combining features
• For linear SVMs
– Early fusion: concatenation the descriptors
– Late fusion: learning weights to combine the classification scores

• Theoreticallyy no clear winner

• In p
practice late fusion g
give better results
– In particular if different modalities are combined
 Multi-class
Multi class SVMs
• Various direct formulations exist
exist, but they are not widely
used in practice. It is more common to obtain multi-class
SVMs by combining two-class
two class SVMs in various ways

• One versus all:

– Training: learn an SVM for each class versus the others
– Testing: apply each SVM to test example and assign to it the
class of the SVM that returns the highest decision value

• One
O versus one:
– Training: learn an SVM for each pair of classes
– Testing: each learned SVM “votes”
votes for a class to assign to the test
example
 Why does SVM learning work?

• Learns foreground and background visual words

foreground words – high weight

background words – low weight

 Illustration

Localization according to visual word probability

Correct − Image: 35 Correct − Image: 37

20 20

40 40

60 60

80 80

100 100

120 120

50 100 150 200 50 100 150 200

Correct − Image: 38 Correct − Image: 39

20 20

40 40

60 60

80 80

100 100

120 120

50 100 150 200 50 100 150 200

foreground word more probable

background word more probable

 Illustration
A linear SVM trained from positive and negative window descriptors

A few of the highest weighted descriptor vector dimensions (= 'PAS + tile')

+ lie on object boundary (= local shape structures common to many training exemplars)
Bag-of-features
Bag of features for image classification
• Excellent results in the presence of background clutter

bikes books building cars people phones trees

Examples for misclassified images

Books- misclassified into faces, faces, buildings

Buildings- misclassified into faces, trees, trees

Cars- misclassified into buildings, phones, phones

 Bag of visual words summary

• Advantages:
– largely unaffected by position and orientation of object in image
– fixed length vector irrespective of number of detections
– veryy successful in classifying
y g images
g according g to the objects
j they
y
contain

• Disadvantages:
– no explicit use of configuration of visual word positions
– no model of the object location
 Evaluation of image classification
• PASCAL VOC [05
[05-12]
12] datasets

• PASCAL VOC 2007

– Training and test dataset available
– Used to report state
state-of-the-art
of the art results
– Collected January 2007 from Flickr
– 500 000 images downloaded and random subset selected
– 20 classes
– Class labels per image + bounding boxes
– 5011 ttraining
i i iimages, 4952 ttestt iimages

• Evaluation measure: average precision

PASCAL 2007 dataset
PASCAL 2007 dataset
Evaluation
 Precision/Recall

• Ranked list for category A :

A, C, B, A, B, C, C, A ; in total four images with category A

 Results for PASCAL 2007
• Winner of PASCAL 2007 [[Marszalek et al.]] : mAP 59.4
– Combination of several different channels (dense + interest
points, SIFT + color descriptors, spatial grids)
– Non-linear
N li SVM with
ith G
Gaussian
i kkernell

• Multiple kernel learning [Yang et al

al. 2009] : mAP 62
62.2
2
– Combination of several features
– Group-based
p MKL approach
pp

• Combining object localization and classification

[Harzallah et al.’09] : mAP 63.5
– Use detection results to improve classification

• Adding objectness boxes [Sanchez at al.’12] : mAP 66.3

 Spatial pyramid matching
• Add spatial information to the bag
bag-of-features
of features

• Perform
P f matching
t hi ini 2D iimage space

[Lazebnik, Schmid & Ponce, CVPR 2006]

 Related work
Similar approaches:
Subblock description [Szummer & Picard, 1997]
SIFT [Lowe, 1999]
GIST [Torralba et al., 2003]

SIFT Gist

Szummer & Picard (1997) Lowe (1999

(1999, 2004) Torralba et al
al. (2003)
Spatial pyramid representation

Locally orderless
representation
i at
several levels of
spatial resolution

level 0
Spatial pyramid representation

Locally orderless
representation
i at
several levels of
spatial resolution

level 0 level 1
Spatial pyramid representation

Locally orderless
representation
i at
several levels of
spatial resolution

level 0 level 1 level 2

 Spatial pyramid matching
• Combination of spatial levels with pyramid match kernel
[Grauman & Darell’05]
• Intersect histograms, more weight to finer grids
 Scene dataset [Labzenik et al.’06]
Coast Forest Mountain Open country Highway Inside city Tall building Street

Suburb Bedroom Kitchen Living room Office

Store Industrial

4385 images
155 categories
c ego es
Scene classification

L Single-level
Single level Pyramid
0(1x1) 72.2±0.6
1(2x2) 77.9±0.6 79.0 ±0.5
2(4x4) 79.4±0.3 81.1 ±0.3
3(8x8) 77.2±0.4 80.7 ±0.3
Retrieval examples
Category classification – CalTech101

L Single-level Pyramid
0(1x1) 41.2±1.2
1(2x2) 55.9±0.9 57.0 ±0.8
2(4x4) 63.6±0.9 64.6 ±0.8
3(8x8) 60 3±0 9
60.3±0.9 64 6 ±0.7
64.6 ±0 7
 Evaluation BoF – spatial
Image classification results on PASCAL
PASCAL’07
07 train/val set

(SH, Lap, MSD) x (SIFT,SIFTC) AP

spatial layout
1 0.53

2x2
3x1

1,2x2,3x1
 Evaluation BoF – spatial
Image classification results on PASCAL
PASCAL’07
07 train/val set

(SH, Lap, MSD) x (SIFT,SIFTC) AP

spatial layout
1 0.53

2x2 0.52
3x1 0.52

1,2x2,3x1 0.54

Spatial layout not dominant for PASCAL’07 dataset

C bi i iimproves average results,
Combination l ii.e., iit iis appropriate
i ffor
some classes
 Evaluation BoF - spatial

Image classification results on PASCAL’07 train/val set

for individual categories
g

1 3x1
Sheep 0.339 0.256
Bird 0.539 0.484
DiningTable 0.455 0.502
Train 0.724 0.745

Results are category

g y dependent!
p
 Combination helps somewhat
 Discussion

• Summary
– Spatial pyramid representation: appearance of local image
patches
t h + coarse global
l b l position
iti iinformation
f ti
– Substantial improvement over bag of features
– Depends on the similarity of image layout

• Recent extensions
– Flexible, object-centered grid
• Shape
p masks [[Marszalek’12]] => additional annotations
– Weakly supervised localization of objects
• [Russakovsky et al.’12]
 Recent extensions

• Efficient Additive Kernels via Explicit Feature Maps

[Perronnin et al.
al.’10,
10, Maji and Berg’09,
Berg 09, A. Vedaldi and Zisserman’10]
Zisserman 10]

• Recently improved aggregation schemes

– Fisher vector [Perronnin & Dance ‘07]
– VLAD descriptor [Jegou, Douze, Schmid, Perez ‘10]
– Supervector [Zhou et al. ‘10]
– Sparse coding [Wang et al. ’10, Boureau et al.’10]

• Improved performance + linear SVM

Fisher vector

 Use a Gaussian Mixture Model as vocabulary

 Statistical measure of the descriptors of the image w.r.t the GMM
 D i ti off likelihood
Derivative lik lih d w.r.t.
t GMM parameterst

GMM parameters:
weight
mean
co-variance (diagonal)

Translated cluster →
large derivative on for this
component

[Perronnin & Dance 07]

Fisher vector

For image retrieval in our experiments:

- only
l ddeviation
i ti wrtt mean, di
dim: K*D [K number
b off Gaussians,
G i D di
dim off d
descriptor]
i ]
- variance does not improve for comparable vector length
Image classification with Fisher vector
• Dense SIFT
• Fisher vector (k=32 to 1024, total dimension from approx.
5000 to 160000)
• Normalization
– square-rooting
– L2 normalization
– [Perronnin’10], [Image categorization using Fisher kernels of non-iid
image models, Cinbis, Verbeek, Schmid, CVPR’12]

• Classification approach
– Linear classifiers
– One versus
ers s rest classifier
Image classification with Fisher vector
• Evaluation on PASCAL VOC’07 linear classifiers with
– Fisher vector
– Sqrt transformation of Fisher vector
– Latent GMM of Fisher vector

• Sqrt transform + latent MOG

models lead to improvement
p

• State-of-the-art performance
obtained
bt i d with
ith lilinear classifier
l ifi
 Evaluation image description
Fisher versus BOF vector + linear classifier on Pascal Voc’07

•Fisher improves over BOF

•Fisher comparable
p to BOF +
non-linear classifier
•Limited gain due to SPM
on PASCAL
•Sqrt helps for Fisher and BOF
•[Chatfield et al
al. 2011]
 Large-scale
Large scale image classification
has 14M images
g from 22k classes

Standard Subsets
– ImageNet
I N t Large
L S
Scale
l Vi
Visuall R
Recognition
iti ChChallenge
ll 2010 (ILSVRC)
• 1000 classes and 1.4M images
– ImageNet10K dataset
• 10184 classes and ~ 9 M images
 Large-scale
Large scale image classification
• Classification approach
– One-versus-rest classifiers
– Stochastic gradient descent (SGD)
– At each step choose a sample at random and update the
parameters using a sample-wise estimate of the regularized risk

• Data reweighting
– Wh
When some classes
l are significantly
i ifi tl more populated
l t d than
th others,
th
rebalancing positive and negative examples
– Empirical
p risk with reweighting
g g

Natural rebalancing, same weight to positive and negatives

 Importance of re-weighting
re weighting

• Plain lines correspond to w-OVR,

d h d one tto u-OVR
dashed OVR

• ß is number of negatives samples

for each positive, β=1 natural
rebalancing

• Results for ILSVRC 2010

• Significant impact on accuracy

• For very high dimensions little impact
 Impact of the image signature size
• Fisher vector (no SP) for varying number of Gaussians +
different classification methods, ILSVRC 2010

• Performance
P f improves
i f higher
for hi h di
dimensional
i l vectors
t
 Experimental results
• Features: dense SIFT,
SIFT reduced to 64 dim with PCA

• Fisher vectors
– 256 Gaussians, using mean and variance
– Spatial pyramid with 4 regions
– Approx. 130K dimensions (4x [2x64x256])
– Normalization: square-rooting and L2 norm

• BOF: dim 1024 + R=4

– 4960 dimensions
– Normalization: square-rooting and L2 norm
Experimental results for ILSVRC 2010

• Features
F t : dense
d SIFT,
SIFT reduced
d d to
t 64 dim
di with
ith PCA
• 256 Gaussian Fisher vector using mean and variance + SP
(3x1) (4x [2x64x256] ~ 130k dim), square-root + L2 norm
• BOF dim=1024 + SP (3x1) (dim 4000), square-root + L2 norm
• Different classification methods
Large-scale
Large scale experiment on ImageNet10k

16.7

Top-1 accuracy

• Significant gain by data re-weighting, even for high-

dimensional Fisher vectors
• w-OVR > u-OVR
• Improves
Impro es oover
er state of the art
art: 6
6.4%
4% [Deng et
et. al] and
WAR [Weston et al.]
Large-scale
Large scale experiment on ImageNet10k
• Illustration of results obtained with w
w-OVR
OVR and 130K
130K-dim
dim
Fisher vectors, ImageNet10K top-1 accuracy
 Conclusion

• Stochastic training: learning with SGD is well

well-suited
suited for
large-scale datasets

• One-versus-rest: a flexible option for large-scale image

classification

• Class imbalance: optimize the imbalance parameter in

one-versus-rest strategy is a must for competitive
p
performance
 Conclusion

• State-of-the-art performance for large-scale image

classification

• Code on
on-line
line available at http://lear
http://lear.inrialpes.fr/software
inrialpes fr/software

• Future work
– Beyond a single representation of the entire image
– Take into account the hierarchical structure