Saliency In Comics
Event details
| Date | 28.08.2019 |
| Hour | 10:00 › 12:00 |
| Speaker | Bahar Aydemir |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Pierre Dillenbourg
Thesis advisor: Prof. Sabine Süsstrunk
Co-examiner: Dr. Mathieu Salzmann
Abstract
Human attention in visual contexts is a widely researched topic which gives clues about the execution process of the human visual system and cognition. Although deep learning methods improved the performances of saliency detection models, the performances appear to be saturated in the last years. Bylinskii et al. explore the commonly missed regions by these models and suggest using high-level concepts to describe them [1]. Moreover, the models that benefit from deep learning methods require large scale datasets. Jiang et al. describe a paradigm to collect large scale saliency data on natural images [2]. This paradigm mimics the human foveal vision by representing the content in different resolutions. Even though salient regions in natural settings are studied by many, there are few approaches for the detection of salient regions in comics images. Thirunarayanan et al, present a method to find and segment important parts on legacy comics by using eye-tracking and generated gaze points [3]. Lastly, we introduce our research plan to detect salient regions on comics by using crowdsourcing and eye-tracking.
Background papers
Where Should Saliency Models Look Next?, by Bylinskii Z., et al.
SALICON: Saliency in Context , by Jiang M., et al.
Creating Segments and Effects on Comics by Clustering Gaze Data , by Thirunarayanan I., et al.
Exam president: Prof. Pierre Dillenbourg
Thesis advisor: Prof. Sabine Süsstrunk
Co-examiner: Dr. Mathieu Salzmann
Abstract
Human attention in visual contexts is a widely researched topic which gives clues about the execution process of the human visual system and cognition. Although deep learning methods improved the performances of saliency detection models, the performances appear to be saturated in the last years. Bylinskii et al. explore the commonly missed regions by these models and suggest using high-level concepts to describe them [1]. Moreover, the models that benefit from deep learning methods require large scale datasets. Jiang et al. describe a paradigm to collect large scale saliency data on natural images [2]. This paradigm mimics the human foveal vision by representing the content in different resolutions. Even though salient regions in natural settings are studied by many, there are few approaches for the detection of salient regions in comics images. Thirunarayanan et al, present a method to find and segment important parts on legacy comics by using eye-tracking and generated gaze points [3]. Lastly, we introduce our research plan to detect salient regions on comics by using crowdsourcing and eye-tracking.
Background papers
Where Should Saliency Models Look Next?, by Bylinskii Z., et al.
SALICON: Saliency in Context , by Jiang M., et al.
Creating Segments and Effects on Comics by Clustering Gaze Data , by Thirunarayanan I., et al.
Practical information
- General public
- Free