Which Saliency Detection Method is the Best to Estimate the Human Attention for Adjective Noun Concepts?

Marco Stricker; Syed Saqib Bukhari; Seyyed Saleh Mozaffari Chanijani; Mohammad Al-Naser; Damian Borth; Andreas Dengel

In: International Conference on Agents and Artificial Intelligence. International Conference on Agents and Artificial Intelligence (ICAART), February 24-26, ISBN 10.5220/0006198901850195, scitepress, 2/2017.


This paper asks the question: how salient is human gaze for Adjective Noun Concepts (a.k.a Adjective Noun Pairs - ANPs)? In an existing work the authors presented the behavior of human gaze attention with respect to ANPs using eye-tracking setup, because such knowledge can help in developing a better sentiment classification system. However, in this work, only very few ANPs, out of thousands, were covered because of time consuming eye-tracking based data gathering mechanism. What if we need to gather the similar knowledge for a large number of ANPs? For example this could be required for designing a better ANP based sentiment classification system. In order to handle that objective automatically and without using an eye-tracking based setup, this work investigated if there are saliency detection methods capable of recreating the human gaze behavior for ANPs. For this purpose, we have examined ten different state-of-the-art saliency detection methods with respect to the ground-truths, which are human gaze pattern themselves over ANPs. We found very interesting and useful results that the Graph-Based Visual Saliency (GBVS) method can better estimate the human-gaze heatmaps over ANPs that are very close to human gaze pattern.

Weitere Links

ICAART_2017_96_CR.pdf (pdf, 10 MB )

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence