formed and meta-data is prepared. Current mobile image re-
trieval systems on the market include SnapTell[3], Kooaba[2],
Google Goggles[1]. However, they all suer from a long lag
between snapping a picture and the response to the query,
typically 10s of seconds. Our system is much faster.
Communication between the camera-phone and the re-
mote server requires a wireless network connection. The
connection is typically EDGE, 3G, or WiFi, with 3G being
the most prevalent. To send data over the wireless network,
one wishes to send as little data as possible. We extract
low bit-rate local descriptors (CHoG) from the query image
[6, 5] and compress their locations using location histogram
coding [13]. We send the signicantly smaller compressed
features instead of the query image, reducing the time spent
in communication.
A large-scale database on the server suers from large
memory usage. Memory swapping operations slow down
the recognition. We compress the inverted index in RAM to
reduce the memory size and prevent memory swapping op-
erations [8]. Complex geometric verication processes also
increase the response time. We use fast geometric re-ranking
to reduce the time for geometric verication [14]. Our sys-
tem achieves image recognition less than one second and can
provide an interactive experience to the user.
Compared to our previous system [12], we have reduced
query data size by using better compression techniques, op-
timized image processing implementation on the phone, and
improved our recognition engine to accommodate a database
of beyond one million items. In this paper, we briey review
our mobile product recognition system in Sec. 2. We report
the processing time of our system and evaluate the benets
of our approach in Sec. 3. In addition to this paper, a video

上一页下一页