Example-Based Example-Based Object Detection in Object Detection in

Images by Images by ComponentsComponentsMohan, Papageorgious and PoggioMohan, Papageorgious and Poggio

IEEE PAMI 2001IEEE PAMI 2001Presented by Jiun-Hung ChenPresented by Jiun-Hung Chen

April 11, 2005April 11, 2005

SummarySummary Goal: Detect objects in static imagesGoal: Detect objects in static images How: Exampled-based person detection How: Exampled-based person detection

framework by componentsframework by components Heads, legs, left arms and right arms detectorsHeads, legs, left arms and right arms detectors Components are present in the proper geometric Components are present in the proper geometric

configurationconfiguration Person detector combines the results of the Person detector combines the results of the

component detectors for person detectioncomponent detectors for person detection Adaptive combination of classifiers (ACC)Adaptive combination of classifiers (ACC) Harr wavelet transform + support vector Harr wavelet transform + support vector

machines (SVM)machines (SVM) Significantly better than a similar full-body Significantly better than a similar full-body

person detectorperson detector

Previous WorkPrevious Work Object detectionObject detection

MModel-basedodel-based IImage invariancemage invariance EExample-basedxample-based

Classifier combination algorithmsClassifier combination algorithms Bagging, Boosting, Voting and so onBagging, Boosting, Voting and so on

Challenges in Person Challenges in Person DetectionDetection

NNonrigid objects, colors, garment typesonrigid objects, colors, garment types RRotated in depth, partially occluded or in motionotated in depth, partially occluded or in motion

System DiagramSystem Diagram

Geometric ConstraintsGeometric Constraints

Harr Wavelet TransformHarr Wavelet Transform

From www.matlab.comFrom www.matlab.com

Support Vector Machines Support Vector Machines (SVM)(SVM)

First, project input data First, project input data nonlinearly and implicitly by nonlinearly and implicitly by kernel functions to a feature kernel functions to a feature spacespace Mercer’s kernels (Polynomial Mercer’s kernels (Polynomial

kernels and Gaussian radial kernels and Gaussian radial basis function kernels)basis function kernels)

Second, find optimal Second, find optimal decision hyperplane in the decision hyperplane in the feature space by maximizing feature space by maximizing soft margins and an upper soft margins and an upper bound of training errorsbound of training errors

The raw output of an SVM The raw output of an SVM classifier is the distance of a classifier is the distance of a data point from the decision data point from the decision hyperplanehyperplane

Training ExamplesTraining Examples

Experimental ResultsExperimental Results

Experimental Results Experimental Results (Cont.)(Cont.)

Experimental Results Experimental Results (Cont.)(Cont.)

Experimental Results Experimental Results (Cont.)(Cont.)

Experimental Results Experimental Results (Cont.)(Cont.)

Learned Geometric Learned Geometric ConstraintsConstraints

ConclusionsConclusions CComponentomponent––based person detectionbased person detection

BBetter than full-body person detectoretter than full-body person detector Hierarchical Classifiers or Adaptive Hierarchical Classifiers or Adaptive

Combination of Classifiers (ACC)Combination of Classifiers (ACC)

Future WorkFuture Work FFace detection: Heisele et al. CVPR’01ace detection: Heisele et al. CVPR’01 Face recognition: Heisele et al. Face recognition: Heisele et al.

CVIU’03CVIU’03 Car detection: Bileschi, Leung and Car detection: Bileschi, Leung and

Rifkin ECCV 04 WorkshopRifkin ECCV 04 Workshop Arbitrary viewpoints?Arbitrary viewpoints?

HHow appearance and geometric ow appearance and geometric configuration changeconfiguration change

QuestionsQuestions LightingLighting Videos Videos

Space-time component based detection, Space-time component based detection, recognition and trackingrecognition and tracking

Other applicationsOther applications InsectInsect

What are meaningful components?What are meaningful components? Object detection/recognition/tracking if Object detection/recognition/tracking if

cameras intrinsic and extrinsic parameters cameras intrinsic and extrinsic parameters may change may change

Thank You! Thank You!