Accurate eye center localization via hierarchical adaptive convolution

Haibin Cai, Bangli Liu, Zhaojie Ju, Serge Thill, Tony Belpaeme, Bram Vanderborght, Honghai Liu

Research output: Chapter in Book/Report/Conference proceedingConference contribution

72 Downloads (Pure)

Abstract

Eye center localization has been an active research topic for decades due to its important biological properties, which indicates human’s visual focus of attention. However, accurate eye center localization still remains challenging due to the high degree appearance variation caused by different kinds of viewing angles, illumination conditions, occlusions and head pose. This paper proposes a hierarchical adaptive convolution method (HAC) to localize the eye center accurately while consuming low computational cost. It mainly utilizes the dramatic illumination changes between the iris and sclera. More specifically, novel hierarchical kernels are designed to convolute the eye images and a differential operation is applied on the adjacent convolution results to generate various response maps. The final eye center is localized by searching the maximum response value among the response maps. Experimental results on several publicly available datasets demonstrate that HAC outperforms the start-of-the-art methods by a large margin. The code is made publicly available at https://github.com/myopengit/HAC.
Original languageEnglish
Title of host publicationProceedings of the 29th British Machine Vision Conference
Subtitle of host publicationBMVC 2018
PublisherBritish Machine Vision Association
Number of pages12
Publication statusPublished - 6 Sep 2018
Event29th British Machine Vision Conference - Northumbria, Newcastle, United Kingdom
Duration: 3 Sep 20186 Sep 2018
http://bmvc2018.org/
http://bmvc2018.org/

Conference

Conference29th British Machine Vision Conference
Abbreviated titleBMVC 2018
Country/TerritoryUnited Kingdom
CityNewcastle
Period3/09/186/09/18
Internet address

Fingerprint

Dive into the research topics of 'Accurate eye center localization via hierarchical adaptive convolution'. Together they form a unique fingerprint.

Cite this