UPC @ MediaEval 2013Social Event Detection (Task 1)

Daniel Manchón-VizueteXavier Giró-i-Nieto

Barcelona, Catalonia19th October 2013

Motivation

Challenge

Challenge

Related work

PhotoTOC[Platt et al, PACRIM 2003]

Approach(a) Temporal sorting by each user independently

Hi, I’m John. Hi, I’m Emily.

Approach(b) Temporal-based oversegmentation in mini-clusters

PhotoTOC[Platt et al, PacRim 2003]

Approach(b) Temporal-based oversegmentation in mini-clusters

Approach(c) Sequential merging of mini-clusters

? tavg(·) avg(·) avg(·)avg(·)

Approach(c) Sequential merging of mini-clusters

Weightedmodalities

● creation (or upload) time● geolocation● textual labels● same user

Approach(c) Sequential merging of mini-clusters

Geolocation (d=haversine)Time stamp (d=L1)

Text labels (d=Jaccard) Same user (d=boolean)

Approach(c) Sequential merging of mini-clusters

Weighting factors (wi)

Time GPS Labels User

Learned weights

0.06 0.28 0.22 0.44

0.08 - 0.30 0.60

Approach(c) Sequential merging of mini-clusters

z-scoreAverage and

deviation learned on pairs of photos within the same training event.

Approach(c) Sequential merging of mini-clusters

phifunction

Approach(c) Sequential merging of mini-clusters

decision threhold

Approach(c) Sequential merging of mini-clusters

Results (required only)F1 NMI Divergence F1

Heuristic weights (*)

0.8798 0.9720 0.8268

Learned weights

0.8833 0.9731 0.8316

(*) wtime=0.2, wgeo=0.2, wtext=0.2, wuser=0.4,

Conclusions● Fast solution due to time-sequential nature.

● Divide and conquer.

● Little gain with this optimisation approach.

● Intuition: Thresholds should be event-dependent.

Thank you MediaEval

SED !

@DocXavi#mediaeval13