Welcome to Ming Zhao's homepage!
 |
MAY GOD BLESS YOU! For God so loved the world that he gave his one and only Son, that whoever believes in him shall not perish but have eternal life. (John 3:16) The Spirit gives life; the flesh counts for nothing. (John 6:63) Therefore, if anyone is in Christ, he is a new creation; the old has gone, the new has come! (2 Corinthians 5:17) |
I'm currently working at Google Inc, Mountain View office, USA. Before that, I worked as a research fellow (postdoc), doing multimedia retrieval and face recognition, in National University of Singapore (NUS), under the supervision of Prof. Chua Tat-Seng, and also worked with Prof. Ramesh Jain and Prof. Terence Sim.
What's new?
Two papers are accepted by CVPR 2010 (details coming soon)
|
Google Landmark: Tour the World
|
|
|
Face Recogntion on Web Videos and Personal Photo Ablums
|
|
Concept Detection/Visual Object Recognition
|
，
TRECVID
2005 (Jul. 2005 -- Sep. 2005)
TRECVID is TREC Video Retrieval Evaluation. It is sponsored by the National
Institute of Standards and Technology (NIST). TRECVID is the most challenging evaluation
for video retrieval in the world. Most famous research groups in video
retrieval participated, such as IBM, CMU, Columbia University.There are four
tasks this year: Shot boundary detection, Low-level feature extraction (camera
motion), High-level feature extraction, Search and Exploring BBC rushes. We
gained the first position in the search task.
，
3D
Face Reconstruction and Animation (May. 2005 -- Now)
3D face reconstruction from image(s) has a wide range of applications, such as
face animation and recognition. The slow speed of the 3D morphable model is due
to the texture mapping. To improve the speed, we only use the shape matching to
recover the 3D shape and use texture mapping to get the texture. However, only
with the shape information, one image is not enough for accurate 3D face
reconstruction. So we propose to use multiple images with the morphable shape
model. First, with the feature points given on the multiple images, the 3D
coordinates of the feature points are estimate by the pose estimation. Then,
frontal and profile 2D morphable shape models are built to estimate the 3D
morphable shape model. These two steps works iteratively to improve the result.
At last, the texture is extracted from multiple images with the pose estimation
from the estimated 3D face.
，
Photo
Album Annotation (Feb. 2005 -- Now)
Home photos are becoming more common place and large quantity of home photos
are available on the Internet. There is a need of efficient techniques to
manage this large collection of photos, some with text annotations but many
without. Basically, we need to identify the following essential attributes in
home photos like the place, time, people. With these attributes, a series of
questions can be asked about photos by time, place, people, and their
combinations.
，
TRECVID
2004 (Aug. 2004 -- Sep. 2004)
TRECVID is TREC Video Retrieval Evaluation. It is sponsored by the National
Institute of Standards and Technology (NIST). TRECVID is the most challenging
evaluation for video retrieval in the world. Most famous research groups in
video retrieval participated, such as IBM, CMU, Columbia University.There are
four tasks this year: Shot boundary detection, Story segmentation,Feature
extraction and Search. We gained the first position in the search task.
，
PersonX
Detection in News Video (Aug. 2004 -- Feb. 2005)
With the development of computer technology, more and more digital videos are
available, which demands more efficient access to video content. Video
retrieval thus becomes a hot research topic in multimedia. To achieve the goal
of video retrieval, it's important to find objects of interest to users in
video. For news video, which is a significant source of video, persons are the
most important objects. Thus finding a specific person, called finding "Person-X",
is essential to understand and retrieve news video. The goal of finding
Person-X is to find the shots where Person-X visually appears.
，
Face
Alignment for Face Recognition (Aug. 2003 ~ Jul. 2004)
I aimed to use face alignment to improve the performance of face
recognition. Although face alignment is very important for high performance
face recognition, existing face recognition systems often use simple alignment
strategies or assume that alignment is done beforehand. I planed to first
improve face alignment algorithms and then combine face alignment with face
recognition.
，
Face
Alignment and Iris Localization (Dec. 2002 ~ Jul. 2003)
This was the work performed when I was a visiting student in Visual Computing
Group of Microsoft Research Asia. Research was focused on iris localization for
iris recognition and face alignment for face recognition under the supervision
of Stan Z.Li.
，
Face
Alignment for 3D Facial Reconstruction (Sep. 2002 ~ Nov. 2002)
The task for face alignment is to accurately locate facial features such as the
eyes, nose, mouth and outline. Accurate extraction of facial features offers
advantages for many applications and is crucial for highly accurate face
recognition and synthesis. We used face alignment for "Real-Time
Realistic-Looking 3D Facial Reconstruction and Interaction by Voice-Driven
Expression Animation", supported by National Natural Science Foundation of
China (60203013).
，
Computer
Aided Medical Imaging Diagnosis. (Dec. 2001 ~ Apr. 2002)
I cooperated with other medical students to develop a system to help medial
imaging diagnosis for the "Science Research Challenge Cup" of
Zhejiang University. This system won the third prize.
，
Content-Based
Video Retrieval and Browsing (Aug. 2001 ~ Jul. 2002)
The goal is to help people to rapidly get desired videos and efficiently grasp
the idea of their contents. We developed a system of video analysis,
segmentation, abstraction, classification, indexing, retrieval and browsing. As
for home video abstraction, we proposed an audio and video combined algorithm
which is especially suitable for home videos.
，
Video
Object Segmentation (Sep. 1999 ~ Jul. 2001)
The goal is to segment semantic video objects from videos. We developed two
techniques: statistical inference based automatic video object segmentation and
hierarchy optical flow based semi-automatic video object segmentation.
Last updated on Feb 25, 2010 PDT.
