A simple presentation of the article: "Cluster-based landmark and event detection for tagged photo collections" on the IEEE MultiMedia magazine.
http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=5611558
11. SURF
SIFT
visual similarity casa mila, la pedrera
tag similarity
co-occurrence
latent semantic indexing
12. step 2: use graph to cluster the photos
1 2
landmark landmark
event
4 3
13. the concept of node structure
neighborhood of node v + node itself = structure of node v
v v v
N(v) v Γ(v)
14. the concept of structural similarity (1)
v
u
Γ(v) ∩ Γ(u)
structural similarity between nodes v and u
Γ(v) Γ(u)
15. the concept of structural similarity (2)
high structural similarity
photo cluster 1
C
A
B
photo cluster 2
low structural similarity
16. # edges
complexity
O (km m) graph-based clustering
average node degree
# dimensions
# clusters
k-means clustering O (I C n D)
# iterations
# nodes
O (n2 log n)
hierarchical agglomerative clustering
22. cluster tag filtering
CLUSTER TAGS
helado tropical barcelona cielos spain field
park güell jaume oller park sclupture el beso
low frequency tags
generic tags
24. 207,750 photos
7,768 users
33,959 unique tags
compare graph-based vs. k-means clustering
user study geospatial coherence
high geospatial
coherence
low geospatial
coherence
25. user study
VISUAL
precision recall κ-statistic
graph-based 1.000 0.110 1.000
k-means 0.806 0.324 0.226
TAG
precision recall κ-statistic
graph-based 0.950 0.182 0.820
k-means 0.848 0.307 0.564
26. geospatial coherence
VISUAL
radius std. deviation
graph-based 357 m 1.18 km
k-means 2.4 km 1.73 km
TAG
graph-based 456 m 1.15 km
k-means 767 m 1.76 km