Database Reference
In-Depth Information
Philbin et al. proposed a soft weighting scheme for object retrieval in large
scale image databases [ 130 ]. This soft-assignment maps high-dimensional SIFT
descriptors to a weighted combination of visualwords, rather than to a single
visualword as hard assignment. The soft-weighting assignment is designed as an
exponential function of the distance to the cluster center. This method allows the
inclusion of features which are lost in the quantization stage. Jégou et al. also
suggested to improve the BoW model by aggregating local descriptors into a
compact short binary coded image representation called Hamming embedding (HM)
[ 124 , 125 ]. At the retrieval stage, a tf-idf based index is built with an integration of
weak geometric consistency verification mechanism to penalize those descriptors
which are not consistent in angle and scale.
4.2.2
Mobile Visual Search
4.2.2.1
Mobile Visual Search in Industry
Due to its potential for practicality, mobile visual search is one of the research areas
drawing extensive attention from both industry and academia. Table 4.1 summarizes
representative mobile visual search applications from industry. Different from the
above mentioned applications, the system described in this chapter is innovative
in terms of an interactive gesture-based (using advanced multi-touch function)
visual search system to help users to specify their visual intent, with a consequent
recommendation based on the visual search results and contextual information. In
this perspective, our system leverages visual search results to formulate a second
query to accomplish task completion on mobile devices, which is significantly
different from existing applications.
Table 4.1 Summary of mobile visual search applications in industry
Application
Features
Techniques
Company
VS a ,OCR b
Goggles
Product, barcode, cover, landmark,
name card, artwork
Google
Bing Vision
Cover, art, text, barcode
VS, OCR
Microsoft
Flow
Cover (CD/DVD/book/video-games),
barcode
VS
Amazon A9 Laboratory
Kooaba
Logos, cover, landmarks
VS
Smart Visuals
Lookthatup
Paintings, posters, labels
VS
LTU Technologies
OCR, AR c
WordLens
Real-time English/Spanish translation
QuestVisual
a Visual search
b Optical character recognition
c Augmented reality
 
Search WWH ::




Custom Search