Visual Phrases and Data Mining

levelsordΔιαχείριση Δεδομένων

20 Νοε 2013 (πριν από 3 χρόνια και 6 μήνες)

84 εμφανίσεις

Ivette Carreras

Haroon

Idrees

Results from
BoVP


Using 5M features for 5K images


Vocabulary size 1K


Measurement : mean Average Precision (
mAP
)


Results from
BoW

18%


Results from
BoVP

1%


Differnces


Vocabulary size is too small for 5M features


5M

50K words


Every phrase is used in retrieval (length 2:6)


Same weight for every length


Transactions are created for every word in each images


Do not discriminate between interesting and not
interesting areas in the image


Van
Gool’s

paper focuses only on certain areas

Radial Transaction configuration

A

A


Rotation invariant


Not
sensible to scale


Unless many new other
words (not part of the
phrase) appear


Transactions are already
mined and sorted

B

C

D

B

Next steps



Increase the vocabulary to 10K and decrease the
number of features


Fix the Visual Phrases code