Face Detection by BoostingObjectiveTesting Phase - Scheme of Viola-Jones face detectorTraining Phase - Training Dataset FeatureFeature PoolFeature PoolTraining Phase - How to Build the Cascaded DetectorExperimental SetupExperimental ResultsExperimental ResultsExperimental ResultsExperimental ResultsAnalysisReferencesThanks!Face Detection by BoostingLiefei XuApril 30, 2009CS559 Machine Learning Course ProjectObjective Detecting Faces in Given ImagesAlgorithm Viola-Jones face detector Use a set of weak classifiers to create a strong classifier.- Weak classifiere.g. a classifier has correct detection rate 99.9% + false positive rate 40%. - Strong classifiere.g. a classifier has correct detection rate 98% + false positive rate 0.01%Testing Phase- Scheme of Viola-Jones face detectorIf we have 10 layers, at each layer correct detection rate is 99.9%, false positive rate is 40%,the final correct detection rate is 99.9%^10=99%, false positive rate is 40%^10=0.01%Training Phase- Training Dataset MIT CBCL dataset All faces are cropped and resized to 19*19 Non-face images use the same image size.FeatureA BFeature = I(White area) – I(Black area) The differences between different features are the location and size of the white and black area.Classifier A: if I(White area) – I(Black area) > θ1then sub-window is a faceClassifier B:if I(White area) – I(Black area) > θ2then sub-window is a faceFeature Pool…………ABCDPrototype…EFeature Pool• number of features for a sub-window size of 19*19Feature Type #A 34,200B 34,200C 21,660D 21,660E 16,200Sum 127,920Training Phase- How to Build the Cascaded Detector At each layer, combine several classifier to create a weak classifier Cascade the weak classifiers at each layer to create a strong classifier Stop when the false positive rate decreases to the requirement. Otherwise, add one more layer.Feature a Feature bclassifier1 classifier2layer1Feature a Feature bclassifier1 classifier2layer2Feature c Feature dclassifier3 classifier4Experimental Setup All experiments were run on Dell XPS 730x desktop, CPU Intel® Core™ i7-920 , 6G memory. Matlab , GML AdaBoost Matlab Toolbox Training dataset - MIT CBCL : ◦ 1stlayer: 1000 faces + 2000 non-faces◦ 2ndlayer: 1000 faces + FP non-faces from 1stlayer◦ 3rdlayer: 1000 faces + FP non-faces from 2ndlayerTesting dataset – MIT CBCL & LFW ◦ 500 faces + 500 non-faces from MIT CBCL testing dataset◦ 1000 faces from Labeled Face in the Wild dataset2 3-layer Detectors◦ Detector 1: 25 – 25 – 63 features◦ Detector 2: 12 – 12 – 20 featuresExperimental Results Training◦ Computing features Computing 127,920 feature values, Over total 3000 (1000 faces + 2000 non-faces) images, Each image is 19*19, costs 33 hours.◦ Selecting featuresLayer # of FeaturesTraining time (in minutes)RecallFalse Positive Rate# of False PositiveDetector 1 Layer 1 25 60.93 99.9% 30.77% 444Layer 2 25 31.62 99.9% 16.19% 192Layer 3 63 58.85 100% 0.00% 0Detector 2 Layer 1 12 33.28 99.2% 48.92% 942Layer 2 12 23.43 99.3% 35.10% 530Layer 3 20 26.83 99% 15.96% 178Layer1 FPLayer2FPLayer3FPExperimental Results Tests on 500 faces+500 non-faces (19*19) from MIT CBCL datasetTesting time(in seconds)Recall FalsePositive RateCombined RecallCombined False Positive RateDetector 1 Layer 1 15 94.8% 11.40% 94.8% 11.40%Layer 2 97.0% 25.27% 92.2% 5.92%Layer 3 95.4% 21.93% 88.0% 3.72%Detector 2 Layer 1 18 97.8% 27.45% 97.8% 27.45%Layer 2 96.8% 29.45% 95.0% 16.08%Layer 3 82.6% 24.50% 79.8% 7.42% Tests on 1000 faces from Labeled Face in the Wild datasetExperimental ResultsScale Testing time (in hour)RecallFalse Positive RateDetector 1 1.1 2.933 83.0% 51.63%1.2 1.624 84.3% 53.50%Detector 2 1.1 2.3 78.4% 63.55%1.2 1.382 80.2% 63.11%Experimental ResultsAnalysis More training data 1000 faces+2000 non-faces vs. 4916 faces+10,000 non-faces Profile image and rotated features Use 24*24 instead of 19*19 faces 24*24 is tested better, includes extra visual information Add more layers 3 layers 113 features vs. 32 layers 4297 features Improve matlab code, decrease detecting timeReferences Paul Viola, Michael Jones, “Robust Real-time Object Detection”, 2ndinternational workshop on statistical and computational theories of vision – modeling, learning, computing, and sampling, 2001 Rainer Lienhart, Alexander Kuranov, Vadim Pisarevsky, “Empirical Analysis of Detecion Cascades of Boosted Classifiers for Rapid Object Detection”, Microprocessor Research Lab, Intel Labs, Technical Report, Dec. 2002 The laboratory of computer graphics at dept. of CS MSU, “GML AdaBoost Matlab Toolbox
View Full Document