Statistical Learning of MultiView Face Detection Microsoft Research

Statistical Learning of Multi-View Face Detection Microsoft Research Asia Stan Li, Long Zhu, Zhen Qiu Zhang, Andrew Blake, Hong Jiang Zhang, Harry Shum Presented by Derek Hoiem

Overview Ø Viola-Jones Ada. Boost Ø Float. Boost Approach Ø Multi-View Face Detection Ø Float. Boost Results Ø Float. Boost vs. Ada. Boost Ø Float. Boost Discussion

Face Detection Overview Ø Evaluate windows at all locations in many scales Classifier Object Non-Object

Viola-Jones Ada. Boost Ø Weak classifiers formed out of simple features Ø In sequential stages, features are selected and weak classifiers trained with emphasis on misclassified examples Ø Integral images and a cascaded classifier allow real-time face detection

Viola-Jones Features For a 24 x 24 image: 190, 800 semi-continuous features Ø Computed in constant time using integral image Ø Weak classifiers consist of filter response threshold Ø Vertical Horizontal On-Off-On Diagonal

Integral Image y = I 8 – I 7– I 6 + I 5+ I 4 – I 3 – I 2 + I 1 I( x 1 , y 1 ) I( x 3, y 3 ) I( x 2, y 2 ) I( x 4, y 4 ) I( x 5, y 5 ) I( x 6, y 6 ) I( x 7, y 7 ) I( x 8, y 8 )

Cascade of Classifiers Input Signal (Image Window) 40% Stage 1 1 Weak Classifier 60% 40% Stage 2 5 Weak Classifiers Class 2 (Non-Face) 60% 99. 999% … 0. 001% Stage N 1200 Weak Classifiers Class 1 (Face) 40%

Viola-Jones Ada. Boost Algorithm Ø Strong classifier formed from weak classifiers: Ø At each stage, new weak classifier chosen to minimize bound on classification error (confidence weighted): Ø This gives the form for our weak classifier:

Viola-Jones Ada. Boost Algorithm

Viola-Jones Ada. Boost Pros and Cons Ø Very fast Ø Moderately high accuracy Ø Simplementation/concept Ø Greedy search through feature space Ø Highly constrained features Ø Very high training time

Float. Boost Weak classifiers formed out of simple features Ø In each stage, the weak classifier that reduces error most is added Ø In each stage, if any previously added classifier contributes to error reduction less than the latest addition, this classifier is removed Ø Result is a smaller feature set with same classification accuracy Ø

MS Float. Boost Features Microsoft For a 20 x 20 image: over 290, 000 features (~500 K ? ) Ø Computed in constant time using integral image Ø Weak classifiers consist of filter response threshold Ø Viola-Jones

Float. Boost Algorithm

Float. Boost Weak Classifiers Can be portrayed as density estimation on single variables using average shifted histograms with weighted examples Ø Each weak classifier is a 2 -bin histogram from weighted examples Ø Weights serve to eliminate overcounting due to dependent variables Ø Strong classifier is a combination of estimated weighted PDFs for selected features Ø

Multi-View Face Detection Head Rotations In-Plane Rotations: -45 to 45 degrees Out of Plane Rotation: -90 to 90 degrees Moderate Nodding

Multi-View Face Detection Detector Pyramid

Multi-View Face Detection Merging Results Frontal Right Side Left Side

Multi-View Face Detection Summary Ø Simple, rectangular features used Ø Float. Boost selects and trains weak classifiers Ø A cascade of strong classifiers makes up the overall detector Ø A coarse-to-fine evaluation is used to efficiently find a broad range of out-ofplane rotated faces

Results: Frontal (MIT+CMU) Float. Boost/Ada. Boost/RBK Schneiderman 20 x 20 images Float. Boost Ø 3000 original faces, 6000 total Ø 100, 000 non-faces Ø Float. Boost vs. Adaboost

Results: MS Adaboost vs. Viola-Jones Adaboost More flexible features Ø Confidence-weighted Ada. Boost Ø Smaller image size Ø

Results: Profile No Quantitative Results!!!

Float. Boost vs. Ada. Boost Ø Float. Boost finds a more potent set of weak classifiers through a less greedy search Ø Float. Boost results in a faster, more accurate classifier Ø Float. Boost requires longer training times (5 times longer)

Float. Boost vs. Ada. Boost 1 Strong Classifier, 4000 objects, 4000 non-objects, 99. 5% fixed detection

Float. Boost: Pros Ø Very Fast Detection (5 fps multi-view) Ø Fairly High Accuracy Ø Simple Implementation

Float. Boost: Cons Ø Very long training time Ø Not highest accuracy Ø Does it work well for non-frontal faces and other objects?