LAPI @ 2015 Retrieving Diverse Social Images Task:

A Pseudo-Relevance Feedback Diversification Perspective

Bogdan Boteanu, Ionuț Mironică, Bogdan Ionescu

LAPI - University ”Politehnica” of Bucharest, 061071, Romania

Email: {bboteanu,imironica,bionescu}@alpha.imag.pub.ro

University POLITEHNICA of Bucharest

HC pseudo-relevance feedback (HC-RF)

pre-filtering of un-relevant images

uses a hierarchical clustering scheme with feedback determined automatically from initial data

diversification achieved by traversing HC image clusters with respect to the Flickr initial ranking

Proposed approach (1)

MediaEval 2015, Wurzen, Germany 1/5

1. Filter optimization

Viola Jones face detector: • Nf – number of faces [0; 3]

• Nr – number of merging rectangles for a face [1; 4]

Blur detector: • Tb – blur threshold [0; 1]

Distance-based filter (GPS coord): • Td – distance threshold [1; 5]

Find best combination (Nf-Nr-Tb-Td) so that P@250 on devset is maxim

Proposed approach (2)

MediaEval 2015, Wurzen, Germany 2/5

2. Selection of positive and negative examples

Proposed approach (3)

MediaEval 2015, Wurzen, Germany 3/5

I 1 I 2 I N

Image Database (Flickr’s rank)

I 3 I N-1

(N’+N”) << N

…

N” (un-relevant) N’ (relevant)

3. HC clustering and pruning

Proposed approach (4)

MediaEval 2015, Wurzen, Germany 4/5

Hierarchical Clustering

cut point

I 1

I 2

Class 1

I N

I N-1

Class k

(un-relevant) I 3 …

4. Diversification

Proposed approach (5)

MediaEval 2015, Wurzen, Germany 5/5

I 1 I 3 … I 4

Class 1 Class 2 Class n

I 9

I 8

I 2

I 7

I 5

I 15

...

...

... 1

2

4

9

3

15 8 7

5

Output

Thank you!

More details coming during the poster session