This document proposes a method for zero-example event detection using concept language models and event-based concept number selection. It aims to augment simple concept terms with additional words and phrases, and exploit the appropriate number of concepts per event. The method creates event detectors using expanded language models and selects an optimal concept number. It then annotates videos with deep concept detectors and retrieves relevant videos using histogram intersection distance. The method is evaluated on the TRECVID MED14TEST dataset and outperforms other zero-example event detection systems.

1. Concept Language Models and Event-based Concept
Number Selection for Zero-example Event Detection
Damianos Galanopoulos1, Foteini Markatopoulou1,2, Vasileios Mezaris1, Ioannis Patras2
1Information Technologies Institute (ITI), CERTH, Thessaloniki, Greece 2Queen Mary University of London, London, UK
Supported by:
Proposed method
Experimental results
Problem and motivation Background
• Zero-example event detection: retrieving, from
a large video collection, videos that are related
to a given event textual description, where
training examples for this event are not given
• Typical solution: transforming both the event
textual description and the available videos into
concept-based representations using simple
concept terms retrieved from a pre-defined
concept pool
• Motivation:
o Augmenting the simple concept terms with
additional words and phrases
o Exploiting the appropriate number of concepts
per event
• TRECVID MED14TEST dataset
• 25.000 videos
• 20 Pre-Specified Events (E021-E040)
• 13.488 semantic concepts
o 12.988 ImageNet concepts
o 500 event-related concepts (EventNet)
• Mean Average Precision (MAP)
A) Parameter selection for ELM, CLM
• (a) Comparison between different types of ELM
• (b) The performance of different quota of under
the curve area (AUC) for the E035 event
C) Comparisons
• MAP between different zero-example event
detection systems
B) Concept number selection
• (a) The overall performance of different quota of
under the curve area (AUC)
• (b) Comparison between different types of CLM
Event detector creation
• (c) The performance for different types of CLM for
the 20 MED14TEST events
Method outline
• Event detector creation: ELM, CLM, concept
number selection
• Video annotation and retrieval: DCNNs and
histogram intersection distance
Video annotation
Event detector
Video concept scores
More detailed and complicated
description compared to the
simple query textual description
that is given regarding the ad-
hoc video search problem
Retrieving complete videos
related to the event kit;
instead of video fragments
(keyframes) related to a
textual query