Date post: | 02-Apr-2018 |
Category: |
Documents |
Upload: | sumeet-saurav |
View: | 230 times |
Download: | 0 times |
of 29
7/27/2019 Viola Jones
1/29
Robust Real-time Face
Detectionby
Paul Viola and Michael Jones, 2002
Presentation by Kostantina Palla & Alfredo Kalaitzis
School of InformaticsUniversity of Edinburgh
February 20, 2009
7/27/2019 Viola Jones
2/29
Overview
Robust very high Detection Rate (True-Positive
Rate) & very low False-Positive Rate always.
Real Time For practical applications at least 2
frames per second must be processed.
Face Detection not recognition. The goal is to
distinguish faces from non-faces (face detection is the
first step in the identification process)
7/27/2019 Viola Jones
3/29
Three goals & a conlcusion
1. Feature Computation: what features? And howcan they be computed as quickly as possible
2. Feature Selection: select the most discriminatingfeatures
3. Real-timeliness: must focus on potentiallypositive areas (that contain faces)
4. Conclusion: presentation of results anddiscussion of detection issues.
How did Viola & Jones deal with these challenges?
7/27/2019 Viola Jones
4/29
1. Feature Computation
The Integral image representation
2. Feature SelectionThe AdaBoost training algorithm
3. Real-timeliness
A cascade of classifiers
Three solutions
7/27/2019 Viola Jones
5/29
Features Can a simple feature (i.e. a value) indicate
the existence of a face?
All faces share some similar properties
The eyes region is darker than theupper-cheeks.
The nose bridge region is brighter thanthe eyes.
That is useful domain knowledge
Need for encoding of Domain Knowledge:
Locat ion - Size:eyes & nose bridgeregion
Value:darker / brighter
Overview| Integral Image | AdaBoost | Cascade
7/27/2019 Viola Jones
6/29
Rectangle features Rectangle features:
Value = (pixels in black area) - (pixels in white area)
Three types: two-, three-, four-rectangles,Viola&Jones used two-rectangle features
For example: the difference in brightnessbetween the white &black rectangles overa specific area
Each feature is related to a speciallocation in the sub-window
Each feature may have any size
Why not pixels instead of features? Features encode domain knowledge
Feature based systems operate faster
Overview | Integral Image | AdaBoost | Cascade
7/27/2019 Viola Jones
7/29
Integral Image Representation(also check back-up slide #1)
Given a detection resolution of 24x24(smallest sub-window), the set ofdifferent rectangle features is
~160,000 !
Need for speed
Introducing Integral ImageRepresentation
Definition: The integral image at
location (x,y), is the sum of thepixels above and to the left of(x,y), inclusive
The Integral image can be computedin a single pass and only once foreach sub-window!
' , '
formal definition:
, ', '
Recursive definition:
, , 1 ,
, 1, ,
x x y y
ii x y i x y
s x y s x y i x y
ii x y ii x y s x y
Overview | Integral Image | AdaBoost | Cascade
y
x
7/27/2019 Viola Jones
8/29
back-up slide #1
Overview | Integral Image | AdaBoost | Cascade
0 1 1 1
1 2 2 3
1 2 1 1
1 3 1 0
IMAGE
0 1 2 3
1 4 7 11
2 7 11 16
3 11 16 21
INTEGRAL IMAGE
7/27/2019 Viola Jones
9/29
7/27/2019 Viola Jones
10/29
Three goals1. Feature Computation: features must be
computed as quickly as possible
2. Feature Selection: select the mostdiscriminating features
3. Real-timeliness: must focus on potentially
positive image areas (that contain faces)
How did Viola & Jones deal with these challenges?
Overview | Integral Image | AdaBoost | Cascade
7/27/2019 Viola Jones
11/29
Feature selection Problem: Too many features
In a sub-window (24x24) there are~160,000 features (all possiblecombinations of orientation, locationand scale of these feature types)
impractical to compute all of them(computationally expensive)
We have to select a subset of relevantfeatures which are informative - tomodel a face Hypothesis: A very small subset of
features can be combined to form aneffective classifier
How? AdaBoost algorithm
Overview | Integral Image| AdaBoost | Cascade
Relevant feature Irrelevant feature
7/27/2019 Viola Jones
12/29
AdaBoost
Stands for Adaptive boost
Constructs a strong classifier as a
linear combination of weighted simpleweak classifiers
Overview | Integral Image| AdaBoost | Cascade
Strong
classifier
Weak classifier
WeightImage
7/27/2019 Viola Jones
13/29
AdaBoost - Characteristics Features as weak classifiers
Each single rectangle feature may be regardedas a simple weak classifier
An iterative algorithmAdaBoost performs a series of trials, each time
selecting a new weak classifier
Weights are being applied over the set ofthe example images
During each iteration, each example/imagereceives a weight determining its importance
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
14/29
AdaBoost - Getting the idea
Given: example images labeled +/-
Initially, all weights set equally
Repeat T times
Step 1: choose the most efficient weak classifier that will be acomponent of the final strong classifier (Problem! Remember the hugenumber of features)
Step 2: Update the weights to emphasize the examples which wereincorrectly classified
This makes the next weak classifier to focus on harder examples
Final (strong) classifier is a weighted combination of the T weak classifiers
Weighted according to their accuracy
Overview | Integral Image| AdaBoost | Cascade
otherwise
xxh
T
t
T
t ttth
02
1)(1
)( 1 1
(pseudo-code at back-up slide #2)
7/27/2019 Viola Jones
15/29
AdaBoostFeature SelectionProblem
On each round, large set of possible weak classifiers (each simpleclassifier consists of a single feature) Which one to choose?
choose the most efficient (the one that best separates theexamples the lowest error)
choice of a classifier corresponds to choice of a feature
At the end, the strong classifier consists of T features
Conclusion
AdaBoost searches for a small number of good classifiers features(feature selection)
adaptively constructs a final strong classifier taking into account thefailures of each one of the chosen weak classifiers (weight appliance)
AdaBoost is used to both select a small set of features and train astrong classifier
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
16/29
7/27/2019 Viola Jones
17/29
Now we have a good face detector We can build a 200-feature
classifier!
Experiments showed that a 200-feature classifier achieves:
95% detection rate 0.14x10-3 FP rate (1 in 14084)
Scans all sub-windows of a384x288 pixel image in 0.7seconds (on Intel PIII 700MHz)
The more the better (?) Gain in classifier performance
Lose in CPU time
Verdict: good & fast, but notenough Competitors achieve close to 1 in
a1.000.000 FP rate!
0.7 sec / frame IS NOT real-time.
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
18/29
Three goals1. Feature Computation: features must be
computed as quickly as possible
2. Feature Selection: select the mostdiscriminating features
3. Real-timeliness: must focus on potentially
positive image areas(that contain faces)
How did Viola & Jones deal with these challenges?
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
19/29
The attentional cascade
On average only 0.01% of all sub-windows are positive (are faces)
Status Quo: equal computation time isspent on all sub-windows
Must spend most time only onpotentially positive sub-windows.
A simple 2-feature classifier canachieve almost 100% detection ratewith 50% FP rate.
That classifier can act as a 1st layer ofa series to filter out most negativewindows
2nd
layer with 10 features can tackleharder negative-windows whichsurvived the 1stlayer, and so on
A cascade of gradually more complexclassifiers achieves even betterdetection rates.
Overview | Integral Image| AdaBoost | Cascade
On average, much fewer
features are computed per
sub-window (i.e. speed x 10)
7/27/2019 Viola Jones
20/29
Training a cascade of classifiers
Overview | Integral Image| AdaBoost | Cascade
Strong classifier definition:
otherwise
xxh
T
tt
T
ttth
02
1)(1
)(11 ,
where )1
log(t
t , 1 tt
t
Keep in mind: Competitors achieved 95% TP rate,10-6 FP rate
These are the goals. Final cascade must do better!
Given the goals, to design a cascade we must choose:
Number of layers in cascade (strong classifiers)
Number of features of each strong classifier (the T in definition)
Threshold of each strong classifier (the in definition)
Optimization problem: Can we find optimum combination?
T
t t12
1
TREMENDOUSLY
DIFFICULT
PROBLEM
7/27/2019 Viola Jones
21/29
A simple framework for cascade training
Overview | Integral Image| AdaBoost | Cascade
Do not despair. Viola & Jones suggested a heuristic algorithm forthe cascade training: (pseudo-code at backup slide # 3) does not guarantee optimality
but produces a effective cascade that meets previous goals
Manual Tweaking: overall training outcome is highly depended on users choices
select fi (Maximum Acceptable False Positive rate / layer)
select di (Minimum Acceptable True Positive rate / layer)
select Ftarget (Target Overall FP rate)
possible repeat trial & error process for a given training set
Until Ftarget is met: Add new layer:
Until fi , di rates are met for this layer Increase feature number & train new strong classifier with AdaBoost
Determine rates of layer on validation set
7/27/2019 Viola Jones
22/29
backup slide #3User selects values forf, the maximum acceptable false positive rate per layer and d,
the minimum acceptable detection rate per layer.
User selects target overall false positive rateFtarget.P= set of positive examplesN= set of negative examplesF
0= 1.0;D
0= 1.0; i = 0
WhileFi >Ftargeti++
ni = 0;Fi=Fi-1
whileFi >fxFi-1oni++
oUsePandNto train a classifier with nifeatures using AdaBoost
oEvaluate current cascaded classifier on validation set to determineFiandDioDecrease threshold for the ith classifier until the current cascaded classifier has
a detection rate of at least dxDi-1 (this also affectsFi)
N= IfFi >Ftargetthen evaluate the current cascaded detector on the set of non-face
images and put any false detections into the setN.
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
23/29
7/27/2019 Viola Jones
24/29
Training
Set(sub-windows)
Integral
Representation
Feature
computation
AdaBoostFeature Selection
Cascade trainerTesting phaseTraining phase
Strong Classifier 1
(cascade stage 1)
Strong Classifier N
(cascade stage N)
Classifier cascade
framework
Strong Classifier 2
(cascade stage 2)
FACE IDENTIFIED
Overview | Integral Image| AdaBoost | Cascade
7/27/2019 Viola Jones
25/29
pros Extremely fast feature computation
Efficient feature selection
Scale and location invariant detector Instead of scaling the image itself (e.g. pyramid-filters), we scale the
features.
Such a generic detection scheme can be trained for detection ofother types of objects (e.g. cars, hands)
and cons Detector is most effective only on frontal images of faces
can hardly cope with 45o face rotation Sensitive to lighting conditions
We might get multiple detections of the same face, due tooverlapping sub-windows.
7/27/2019 Viola Jones
26/29
Results(detailed results at back-up slide #4)
7/27/2019 Viola Jones
27/29
Results (Cont.)
7/27/2019 Viola Jones
28/29
Viola & Jones prepared their final Detector cascade: 38 layers, 6060 total features included
1st classifier- layer, 2-features 50% FP rate, 99.9% TP rate
2nd classifier- layer, 10-features 20% FP rate, 99.9% TP rate
next 2 layers 25-features each, next 3 layers 50-features each
and so on
Tested on the MIT+MCU test set
a 384x288 pixel image on an PC (dated 2001) took about 0.067seconds
Detector 10 31 50 65 78 95 167 422
Viola-Jones 76.1% 88.4% 91.4% 92.0% 92.1% 92.9% 93.9% 94.1%
Rowley-Baluja-Kanade 83.2% 86.0% - - 89.2% 89.2% 90.1% 89.9%
Schneiderman-Kanade - - - 94.4% - - - -
Roth-Yang-Ajuha - - - - - - - -
False detections
Detection rates for various numbers of false positives on the MIT+MCU test set containing 130images and 507 faces (Viola & Jones 2002)
backup slide #4
7/27/2019 Viola Jones
29/29
Thank you for listening!