Face detection aims to find all faces in an image or
video sequence in a sense that locates their positions and gives
their approximate sizes and poses as well.
task can be divided into two category according to pose changes of
off-image-plane rotation, that is, frontal face detection and
multi-view face detection. In addition taking in-image-plane
rotation into consideration, it can be further extended to rotation
invariant frontal face detection or rotation invariant multi-view face
Face alignment is a
procedure to get the shape of a human face. Here, the shape
means all the landmarks used to represent a face.
Face alignment can make the
face distribution statistically more compact and provide a good base
for face recognition, face modeling and synthesis.
Face Pose Estimation
Face pose classification as a simplified case of face pose estimation is very
important in many face processing work including face modeling and face recognition.
In general, face pose may represent human physical state, such as sleepiness and concentration.
Therefore, it is very important in many real-life applications, such as monitoring attentiveness
of drivers or automating camera management.
We extract pose related information by pose apperance models, which are built by statistical learning
over a very large training set. Given two eye-centers and mouth center, each face region is warped
to all potential pose models. The face pose can be estimated by maximizing the fitness of the face
to the model.
Face & Head Tracking
Visual cues designed for general object tracking problem
hardly suffice for robust head tracking under diverse circumstances, making it a necessity to utilize higher level information
which is object-specific. To this end we use the vector-boosted multi-view
face detector as the “face cue” in addition to other general
visual cues targeting the entire head. Data fusion is done by an extended particle filter which supports
multiple distinct yet interrelated state vectors (referring to face and head
in our tracking context). Furthermore, pose information provided by the
face cue is exploited to help achieve improved accuracy and efficiency in
the fusion. Experiments show that our algorithm is robust against
target position, size and pose change as well as unfavorable conditions
such as occlusion, poor illumination and cluttered background.
Face Organ State Estimation
and Face Organ Contour Extraction
Face organ state estimation aims to recognize the state of eye and mouth (close or open),
with a real value indicate the degree of openness. we have explored
classical methods in machine learning area such as LDA, Bayes inference,
SVM, Adaboost, etc. On the other hand, Face organ contour extraction aims
to extract the contour of eye and mouth. Our extraction algorithm is implemented
under the framework of Active Shape Model.
recognition aims to identify or verify one or more persons in
an image or video using a stored database of faces. We
focus on face recognition in uncontrolled environments, such as
faces in consumer image collections and faces in the news or
teleplay videos. We are developing statistical as well as
instance-based methods to improve recognition accuracy. Also,
we are interested in developing improved models for better
performance on very large galleries. Moreover, we are trying to
develop better face representations to deal with
variations of illumination and expression.
Face Demographic Classification including
Gender, Age, Ethnicity, etc.
classification aims to tell gender, ethnic and age information
of face images automatically. We are developing demographic
information image filters to extract gender, ethnic and age
information from consumer images.
Face Indexing and Retrieval
Face index and retrieval
try to sort and retrieve face images and videos by their face content
automatically. We are developing tools for the managements of image
or video collections such as digital family photo albums or home
videos, teleplays that mainly involve only a dozen of people or