Face Recognition and Retrieval in Video

29
Face Recognition and Retrieval in Video Basic concept of Face Recog. & retrieval And their basic methods. C.S.E. Kwon Min Hyuk

description

Face Recognition and Retrieval in Video. Basic concept of Face Recog . & retrieval And their basic methods. C.S.E. Kwon Min Hyuk. True? False?. Q1 : in recently, face recognition researches focus on video-based rather than still image-based (O / X) Q2 : There is three approaches; (O/ X) - PowerPoint PPT Presentation

Transcript of Face Recognition and Retrieval in Video

Page 1: Face Recognition  and Retrieval in Video

Face Recognition and Retrieval in VideoBasic concept of Face Recog. & retrieval And their basic methods.

C.S.E.Kwon Min Hyuk

Page 2: Face Recognition  and Retrieval in Video

True? False?Q1 : in recently, face recognition

researches focus on video-based rather than still image-based (O / X)

Q2 : There is three approaches; (O/ X)◦1. key-frame based◦2. Temporal model based◦3. image set based

Face variation and expression make face recognition difficult. (O/X)

Page 3: Face Recognition  and Retrieval in Video

AnswersAll of statement is true.

Page 4: Face Recognition  and Retrieval in Video

Intro Q1 : Why do we need Face recogni-

tion system?◦ Increasing request to search specific peo-

ple related video contents

◦ Can be applied at security, human-com-puter inter action etc.

Page 5: Face Recognition  and Retrieval in Video

IntroQ2: What is recent trend of approach to

Face recog.?

◦ Traditionally , focused on Still image-based appr.

◦ Recently , focused on Video-based appr.

◦We can extract more information from video than that of still image.

Page 6: Face Recognition  and Retrieval in Video

General steps for Face recognitionWhere is face located in video

frame?◦ We should look for which part of the frame

is face. Face detecting and Tracking.

Recognizing face◦There is some basic approach for

Face Recog. Key fame-based approach Temporal Model-based approach Image set-based approach

Page 7: Face Recognition  and Retrieval in Video

Face detectionUsing statistical geometric model

◦ From the frame Extract appearance fea-tures such as edge, intensity, color(histogram)

To evolve the face detector by using machine learning tech.◦Adaboost◦Neural Network◦Support Vector Machine

Page 8: Face Recognition  and Retrieval in Video

Face trackingFace detection’s limit.

◦It detect only frontal or near frontal view.

Tracking face is needed to handle large head motions.

Face tracking◦Difficulties◦Method to solve

Page 9: Face Recognition  and Retrieval in Video

Face trackingDifficulty 1 : There is

◦Face appearance variation◦3D motion◦Background change

Method to solve: face online boosting◦Using tracked images in previous

frames.◦Applying current result to tracking

seq. for next frame. (real-time updat-ing feedback)

example in next slide

Page 10: Face Recognition  and Retrieval in Video

Example for online boosting al-gorithm.

So-called “adaptive tracker.”

Page 11: Face Recognition  and Retrieval in Video

Face tracking(con’d)Difficulty 2: The adaptive tracker

can adapt to non-targets.

Method to solve : add basal appear-ance of target ◦ Teaching the tracker about some basal ap-

pearance of the target. ◦ Basal appearance : Image set of various

target condition (face expression, pose etc)

Page 12: Face Recognition  and Retrieval in Video

After face detecting and trackingWe can determine the part of

frame where face is located.

Now, we can get to face recogni-tion.

Page 13: Face Recognition  and Retrieval in Video

Face RecognitionBasic steps for Face Recog.

◦1. get weak evidence in individual frame.

◦2. collect that evidence over time.◦3. lead(determine) reliable result.

Three approaches◦1. key-frame approach◦2. temporal model-based approach◦3. image set-based approach

Page 14: Face Recognition  and Retrieval in Video

Key-frame based ap-proachTreat each video as a collection

of images. Basic steps of the approach.

◦1. input data(still images, video)◦2. from data, extract images of the

target. Extracted images are called key-frames

or examplars.◦3. matching them with all or subset

of other video sequence(where the target is).

Page 15: Face Recognition  and Retrieval in Video

Key-frame based ap-proach(con’d)How can we get some ‘good’ key-

frame from input data?◦By image-based recognition◦In each frame, probe the nose and

eyes’ triangular structure.

If it is in the frame, then face recognition is performed. And key-frame is extracted.

Page 16: Face Recognition  and Retrieval in Video

Key-frame based ap-proach(con’d)

◦Applying K-Means clustering Cluster means a group whose elements

have some common property. This algorithm is grouping some data ob-

servations into one of cluster which has nearest mean.

Page 17: Face Recognition  and Retrieval in Video

Key-frame based ap-proach(con’d)Other algorithms

◦ Isomap algorithm◦ Combination of majority and probabilistic

voting.◦ And so on. (I’ll skip the details.)

Finally, all or subset of video sequence will be compared(matched) with extracted ‘good’ key-frame to determine recognition.

Page 18: Face Recognition  and Retrieval in Video

Temporal Model Based approachTo handle face dynamics

◦Ex: face expression(non-rigid)or head movement(rigid)

Using temporal sequence(continuous coherent)◦Ex> Using whole sequence of chang-

ing face dynamics as a image set.

Page 19: Face Recognition  and Retrieval in Video

Temporal Model Based approach(con’d)Basic methods

◦Matching the face Trajectory. Trajectory means the moving face’s

path(orbit) through in surfaces. Two(model and object) trajectory dis-

tance accumulates recognition evidence over time.

Page 20: Face Recognition  and Retrieval in Video

Temporal Model Based approach(con’d)Other method

◦Trained statistical face model Using density estimation.

◦Probabilistic approach Using time-series state space variables

◦Hidden Markov Model Fusing pose and person-discriminant fea-

tures.

I’ll skip all of details.

Page 21: Face Recognition  and Retrieval in Video

Image set based approachThis approach uses

◦both image collected over consecutive time

(similar with temporal image set)◦And independent still image set

(similar with key-frame) Combination of both temporal and

key-frame based approaches.Two major approaches.

◦Statistical modal-based◦Mutual subspace-based

Page 22: Face Recognition  and Retrieval in Video

Image set based ap-proach(con’d)Image set classification

◦Non-parametric sample based Compare representative images of each

image sets◦Parametric model-based

In terms of probabilistic, compare two distributionsof each image set.

Page 23: Face Recognition  and Retrieval in Video

Image set based ap-proach(con’d)Statistical Model-based

◦To determine recognition, consider similarity of two manifolds manifold is large group (more than clus-

ter)which contains several cluster.

Drawback ◦Need to solve the difficult parameter

estimation problem.

Page 24: Face Recognition  and Retrieval in Video

Image set based ap-proach(con’d)Mutual Subspace-Based

Model(MSM)◦To determine Similarity between im-

age sets measure by the smallest principal angles

between subspaces. CMSM is expansion of MSM

◦Assume more constraints.

Page 25: Face Recognition  and Retrieval in Video

Face RetrievalIt is difficult to recognize face in the

uncontrolled condition like face dy-namics , light intensity, hair styles

◦There is two applications1. Person Retrieval 2. Cast listing

Page 26: Face Recognition  and Retrieval in Video

Person RetrievalFace recognition tech. are applied.Basic method

◦Using head model (with multiple tex-ture map)

◦Step1. rendering(extract or generate) face images

◦Step2. identifying target face.◦Step3. updating the texture map of

the model.

Page 27: Face Recognition  and Retrieval in Video

Cast listing (cast : actors or charac-ters in film)Automatic cast listing is interest-

ing problem.Based on face recognition

◦Because face is repeatable cue in the film

Using image clustering method

For accuracy, Treat clothing ap-pearance additional cues for clustering.

Page 28: Face Recognition  and Retrieval in Video

Challenges and Future di-rectionDatabases.

◦Constructed in lab. enviro. Not a real world.

◦Limited face appearance of variation.Low-quality Video data

◦Exist lots of noise hard to filter out.Computational Cost

◦Face recognition requires quite high power devices.

Page 29: Face Recognition  and Retrieval in Video

ConclusionFace recog. can be applied in var-

ious area.Face detecting and Tracking.Three general Methods

◦Key-frame based◦Temporal model based◦Image set based

Person Retrieval and Cast listingChallenges to evolve Face recog.

system.