Database Reference
In-Depth Information
The quadkey method is adopted from the Bing Maps Tile System.
1
It converts
the GPS coordinates to a hashing-based representation for fast search and retrieval.
We present an example in Fig.
4.6
to walk through the steps of conversion from
the WGS-84 GPS to a quadruple tiles code. We encode the GPS to a 23 digits
number with the ground resolution of possible 0.02 m accuracy. The formulation
of this distance is computed by the Quadkeys representation. GPS context from
mobile sensor is collected first. The standard WGS-84 is encoded to the quadkey
representation. In the illustration, pictures of the same landmark (the Brussels town
hall) with both the front and the back façades are taken. These two photos have
different WGS-84 information, which have 10 out of 15 quadkey digits identical
after Bing Maps projection. In other words, the hamming distance between these
two codes is 5, which is calculated using tables to approximate a ground distance of
about 305 m.
This section uses a context-aware mobile visual search based on the BoW model
and the hierarchical visual vocabulary tree. Contextual GPS information is also used
in filtering the visual search result. In the next section, an implementation named
TapTell
is presented based on the CVT algorithm introduced.
TapTell
is able to
achieve social activity recommendations through mobile visual searches.
4.3
Mobile Visual Search System for Social Activities Using
Query Image Contextual Model
TapTell
is a system that utilizes visual query input through an advanced multi-
touch mobile platform and rich context to enable interactive visual search and
contextual recommendation. Different from other mobile visual searches,
TapTell
explores users individual intent and their motivation in providing a visual query
with specified ROI. By understanding such intent, associated social activities can be
recommended to users. Existing work has predominantly focused on understanding
the intent expressed by text (or the text recognized from a piece of voice). For
example, previous research attempts to estimate user's search intent by detecting
meaningful entities from a textual query [
131
,
140
]. However, typing takes time
and can be cumbersome on the phone, and thus in some cases, not convenient in
expressing user intent. An alternative is to leverage speech recognition techniques
to support voice as an input. For example, popular mobile search engines enable
a voice-to-search mode.
2
,
3
Siri is one of the most popular applications that further
structure a piece of speech to a set of entities.
4
However, text as an expression of
Search WWH ::
Custom Search