Year: 2023
Type: article
Abstract: We present a novel method for deep image saliency prediction\nthat leverages a cognitive model of visual attention as an inductive bias. This is in stark contrast to recent purely data-driven\nmodels that have achieved performance improvements mainly\nby increased model capacity, resulting in high computational\ncosts and the need for large scale, domain specific training data.\nWe demonstrate that by leveraging a cognitive model of visual\nattention, our method achieves competitive performance to the\nstate-of-the-art across several benchmark natural image datasets while only\nrequiring a third of the parameters. Furthermore, we set\nthe new state of the art for saliency prediction on information\nvisualizations, demonstrating the effectiveness of our approach\nfor cross-domain generalization.We further provide large-scale\ncognitively plausible synthetic gaze data on corresponding images in the full MSCOCO and FigureQA datasets, which we\nused for pre-training. These results are highly promising and\nunderline the significant potential of bridging between firstprinciple cognitive and data-driven models for computer vision\ntasks, potentially also beyond saliency prediction, and even\nvisual attention.
Authors: Sood, Ekta, Shi, Lei, Bortoletto, Matteo, Wang, Yao, Müller, Philipp, Bulling, Andreas
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7019805962
