Video-Lectures over Internet - E-Infrastructures and Technologies for Lifelong Learning - page 263

The cost analysis becomes then very simple.
Costs concern the infrastructure (acquisition sta-
tion and distribution station) plus running costs.
The first one can be very low, as an acquisition
station, with some of the free software solutions,
costs less than 2000 Euro: it only requires a laptop,
a digital camera and a radio-microphone (which is
an essential element, as guaranteeing good audio
quality is a crucial aspect). The distribution sta-
tion - with a terabyte hard disk costing by now
as little as 100 Euro - costs as much as a web
server: 1000 Euros can be more than enough. Of
course, if one foresees many simultaneous users,
a streaming server may be a better choice - in
which case another 1000 Euros might be needed
for the software, even though there exists open
source solutions, such as VLC/VLS 12 or the Darwin
Streaming server 13 . In all cases, the infrastructural
cost should not be a major hurdle.
Running costs may be an issue. In the best
case, a fully automated system will need a peri-
odic supervision, as most services do - but this
would require a very limited annual workload.
Most systems can perform all the needed post-
processing - up to the creation and maintenance
of a dedicated web site in an automatic manner.
The acquisition phase can be automatic, in which
case the cost would be close to zero, or manual. In
the second case the costs can be significant, but
much lower than indicated by Rowe. A dedicated
person could probably collect at least 1000 hours
per year, and would not need a sophisticated techni-
cal specialization, as using most systems is rather
straightforward. Our estimate of the acquisition
cost is hence at most of the order of 25 Euro per
hour, and it can be cut (at least) by two by asking
a student to do simple operations during a lecture
(the cognitive overhead can be minimal, so that
the student can actually follow the lecture while
performing those operations).
On the other hand, a non-fully automated solu-
tion brings some advantages that may be worth
the cost. During acquisition there are essentially
three actions that might need assistance: starting/

pausing/stopping the recording, manipulating the
camera (slewing, zooming, panning), synchroniz-
ing the slides. Among these, camera manipulation
can sensibly increase the quality of video-lectures
- especially if the teacher uses the blackboard.
Not surprisingly, research has been addressing the
problem of creating a fully automatic acquisition
system using a wide variety of approaches, as we
shall discuss in the next section.

RESEARCH ISSUES

Detecting Slide Transitions

The first system able to detect slide transitions
we are aware of was proposed by Mukhopad-
hyay (1999). During the recording phase, a syn-
chronization tone was added to the audio track.
During post-processing all the available images
(e.g. slides) were matched with the portion of the
video showing the projection screen, and the best
fit was chosen.
During the last decade several papers dealt
with computer vision-based, statistical techniques
to attack the same problem, trying to offer bet-
ter solution. The typical pre-processing involves
segmenting the lecturer from the background,
localising the projection area, finding the slide
transitions and matching the observed slides in
the videos to their electronic versions. Reviews
of these attempt can be found in the papers by
Gigonzac et al (2007) and De Lucia et al. (2008).
Generally such approach has the drawback that
the camera should be fixed (to facilitate the lo-
calization of the screen), and most importantly
that the projection screen should to be part of the
video image (which in our view severely limits
the usefulness of the video itself).
Chen and Heng (2003) used an automated
speech recognizer and matched words from the
speech with words from the slide to identify
slide transitions, while Wang et al. (2007) anal-