Domain classifier-based transfer learning for visual attention prediction

Abstract

Benefitting from machine learning techniques based on deep neural networks, data-driven saliency has achieved significant success over the past few decades. However, existing data-hungry models for saliency prediction require large-scale datasets to be trained. Although some studies based on the transfer learning strategy have managed to acquire sufficient information from the limited samples of the target domain, obtaining saliency maps for the transfer process from one image category to another still remains a challenge. To solve this problem, we propose a domain classifier paradigm-based adaptation method for saliency prediction. The method provides sufficient information by classifying the domain from which the data sample originated. Specifically, only a few target domain samples are used in our few-shot transfer learning paradigm, and the prediction results are compared with those obtained through state-of-the-art methods (such as the fine-tuned transfer strategy). To the best of our knowledge, the proposed transfer framework is the first work that conducts saliency prediction while taking the domain adaptation of different image categories into consideration. Comprehensive experiments are conducted on various image category pairs for source and target domains. The experimental results show that our proposed approach achieves a significant performance improvement with respect to conventional transfer learning approaches.