Please answer three out of the first four questions. The
answer to each question shall be limited to three pages in WORD
format with 11pt font size.
Problem 1: Image Search Systems (40 pts)
Please read this survey paper
on Image Searches
and answer the following two questions:
Enumerate the major steps of building an image search engine.
Summarize the major features employed by state-of-art image
search engines.
Problem 2: Interactive Media Servers (40 pts)
Traditional disk IOs are non-preemptible in nature. Specifically,
after an IO is issued, the IO cannot be stop in the
middle of accessing the target disk. Read the Virtual IO paper and answer the following two
questions:
Summarize the three approaches employed by the paper to enable
IO preemption.
Explain why the disk throughput is not affected when
a large data transfer is divided into
a number of small disk accesses.
Problem 3: Query-concept Learning (40 pts)
The SVMActive paper proposes using SVMs and Active Learning to find relevant images in
a search session.
Describe in details the principal steps and ideas of SVMActive.
Enumerate two pros and two cons of the approach.
Problem 4: Similarity (40 pts)
Read the DPF paper and answer the following questions.
Explain the two major problems of measuring ismilarity
in a very high-dimensional space.
How and why DPF works.
What are the shortcomings of DPF, name at least two.
Problem 5: Machine Learning (80 pts)
A quiz will be given during the lecture of 5/21.
Please consult the lecture note
at http://www-db.stanford.edu/~echang/DASFAA-tutorial-handout.pdf