Game Development Reference
In-Depth Information
for certain images). Thus, for future use of the game, the image selection heuristics
responsible for game round setup, would have to respect the past occurrences of
images and select them again only after longer time periods.
5.5 Discussion
To address the challenge of lack of multimedia metadata (particularly for personal
repositories), we devised semantics acquisition games of PexAce and PexAce-
Personal. With their game log processing procedures, we presented a working
approach for acquisition of descriptive image tags. After the controlled and uncon-
trolled deployment and experiments with the game, we can conclude:
Our approach is capable to deliver valid metadata (tags) for general domain of
images.
When used over personal image collections, the approach retains its levels of
tag correctness, even when only very low number of players is participating on
gameplay over the same image collection.
Furthermore, the game yields desired types of metadata needed for personal col-
lections such as person names, places and events.
Although the overall precision of the method is good, there are several secondary
issues, solution of which would improve the output of the game (qualitatively and
quantitatively).
As one source of bias, we identified the automated translation of annotations to
a single language. In particular, this was a problem for shorter annotations, where
the automated translation service could not exploit the word context in case of
homonyms. Moreover, in our particular experiments, where most of the annota-
tions were done in Slovak language, more bias was introduced by the lack of proper
accents in texts, which may significantly change the meaning of individual words.
The solution for this may be to restrict the game language (and annotations) to Eng-
lish only, however, we have concerns about the possible distraction of some players.
As a supportingmechanics to this, may be the auto-complete text typing functionality
(e.g. as in search engines).
This would proactively suggest English words or even concepts (if the auto-
complete box was backed by a domain model) and reduce the bias. A downside of
such approach though, would be probable inefficiency of “general English” dictio-
nary in “personal” version of the method (as many player-specific words would be
used in the game). On the other hand, a specialized gazetteer might prove useful here
(e.g. a list of player's friends, some of which could be expected in the images, could
be acquired through player's social network identity).
Another feature of themethod that can be improved is the player “work” allocation,
i.e. what image is assigned to which game, resp. its effectiveness. In other words for
a constant number of player inputs, we want to maximize the output of the game, i.e.
the number of correct tags. The factor that influences this effectiveness ratio is the
Search WWH ::
Custom Search