Thursday, 29 Jun 2017
You are here:

Subscribe to our Newsletter

How many eyes has a typical person? (ex: 1)
Name:
Email:

Twitter Feed

EMail Print

Event 

Title:
Prof. Touradj Ebrahimi (EPFL) visiting Prof. Jean-Luc Dugelay (Eurecom)
When:
27.03.2015 - 28.03.2015
Category:
Internal Visits

Description

The purpose of this visit was the continuation of the collaboration between Eurecom and EPFL regarding mini-drone project and extend it towards gender recognition using automatic video analytics and crowdsourcing.
Gender recognition is an important topic in video surveillance and can invade privacy of individuals. Therefore, it is important to know the state of the art of gender recognition and assess their performance on realistic content. It is equally important to know how counter measures such as visual privacy filters affect the performance of gender recognition both when observed by humans and analysed by computer vision algorithms.
In this context, a large number of content with full face and body of several hundred individuals were collected. A deep learning algorithm was developed in order to recognize the gender of the images, which were labelled in terms of the gender of the individuals in them. Additionally, a number of privacy filters developed in the framework of VideoSense were applied to the images and both originals and filtered versions were submitted to the gender recognition algorithm developed. In parallel, a crowdsourcing campaign was set up on Microworkers platform and subjects were asked to manually label the gender of individuals in images presented to them, with the latter being either the original or a filtered version. Care was taken into account in order to ensure that no individual was able to see the same content twice in order to avoid the memory effect.
Results show that in many cases, the performance of a deep learning is similar or at times even better than human subjects in gender recognition not only when observing original materials but laos visually filtered for the purpose of gender recognition.
The results obtained were submitted to a conference and will be published at the upcoming SPIE Applications of Digital Image Processing in San Diego, CA, USA in August 2015.