Face Alignment at 3000fps via Regressing Local Binary Features

Face Alignment at 3000fps via Regressing Local Binary FeaturesCVPR14Shaoqing Ren, Xudong Cao, Yichen Wei, Jian Sun

2015.10.02

Presented by Sung Sil Kim1What is face alignment?

Face alignment = landmark detection2Challenges

High accuracy speed 3Traditional ApproachesActive Shape Model (ASM) Continuously track features that form shape of an objectSensitive to noise, sample(person) variation

Active Appearance Model (AAM)Sensitive to initializationHighly discriminative feature needed (fragile to appearance change)Rough estimation

Regression-based method4Cascade Shape Regression Framework

Initial face shape (random training shape) ( groundtruth )

The shape regression approach predicts facial shape S in a cascaded manner [12, 5, 4, 32, 3]. Beginning with an initial shape S0, S is progressively refined by estimating a shape increment S stage-by-stage. In a generic form, a shape increment St at stage t is regressed as: St = Wtt I, St1 , (1)where I is the input image, St1 is the shape from the previous stage, t is a feature mapping function, and Wt is a linear regression matrix.

5Issues with cascade shape regressionPractical issueUsing entire face region as training input -> extremely large feature pool-> unaffordable training costGeneralization issueLarge feature pool has many noisy features->cause overfitting->hurt performance in testing

Approach overviewTree induced local binary featuresLearned from data pixel difference vs. SIFT on landmarksEfficient training/testing3000 FPS on desktop

SIFT