Introduction

3D Face Representation

Face Alignment Databases

Experimental Results

Multiple Constraints

( )1 2

1 2

...:,

...N

N

u u us

v v væ ö

= = +ç ÷è ø

U RA d t

exp1 2

1 2 0 exp exp1 1

1 2

...

... p p

...

idN NN

i i i iN id id

i iN

x x xy y yz z z = =

æ öç ÷= = + +ç ÷ç ÷è ø

å åA A A A

( ), , , , ,T

x ys t ta b g=mexp,T T

idé ù= ë ûp p p

3D scansw. labels

2D image w. landmarks

Training

Databases Pose range Landmark #

Images #

300W Near-frontal 68 3148

AFLW-LFPA All poses 34 3901

Caltech10K Near-frontal 4 10524

300W-LP All poses 68 96268

COFW Near-frontal 29 1007

The main contributions of proposed method:

• Define a new problem of dense face alignment and predict dense shape of the face.

• DeFA model can adopt multiple constrains and leverage multiple datasets.

• Best performance on challenging large-pose face alignment.

Ø Evaluation on large-pose face alignmentsTest on three challenging large-pose datasets

2D landmarks U are represented as pair of projection matrix parameter m and 3D shape parameter p.

Ø SIFT Pairing Constraint (SPC)

first two rowsof rotation matrix

index vector of semantically meaningful

3D vertexes

scalerotationangles translations

Proposed DeFA Model

Testing

Databases Pose range Landmark #

Images #

300W Near-frontal 68 689

AFLW-LFPA all poses 34 1299

AFLW2000-3D all poses 68 2000

IJB-A all poses 3 25795

LFW Near-frontal 0 34356

Ø Loss Function

Ø Parameter Constraint (PC)

Ø Landmark Fitting Constraint (LFC)

Ø Contour Fitting Constraint (CFC)

Stage 1

Stage 2

Stage 3

PC300W-LP

300W-LP Caltech10k COFW

AFLW-LFPA

300W

PC

LFC

LFC

SPC

CFC

Training Procedures

Ø Evaluation on medium-pose face alignments