Project Abstract

Visual Saliency often described as regions and objects within an image that is initially the most viewed upon inspection is a key field in computer vision. Salient object detection describes the abilities of deep learning models to detect these regions of saliency within images. The importance of saliency object detection resonates in various everyday solutions, such as video compression, image labelling, app development, autonomous vehicles and many more. With recent development in technology deep learning models have further undergone progress to deal with complex image analysis to understand and represent saliency better. The limitations presented recently by these models is something we explore as part of this project. A rather new approach within the top-down higher level thinking is a concept that relates saliency with visual relationships, a construct that exists in nature but is quite hard to be defined and implemented by deep learning models. To help this we build a dataset from ground up using a state of the art eye-tracker the “tobii-x3-120” the dataset is used to define this relation often called the “Subject-relation-object” triplet which defines that subjects that are related to certain objects also form part of the salient region. A new dataset through eye-tracking experiments are built and validated on current existing state of the art deep learning models. With this new dataset new models can be built resulting in newer and higher models of thinking within the space of salient object detection.

Keywords: Computer vision, Deep learning, Salient object detection

 Conference Details

Session: Poster Session A at Poster Stand 80

Location: Sir Stanley Clarke Auditorium at Tuesday 7th 13:30 – 17:00

Markers: Gary Tam, Betsy Dayana Marcela Chaparro Rico

Course: BSc Computer Science, 3rd Year

Future Plans: I’m continuing studies