Few-shot personalized saliency prediction using meta-learning

摘要

Personalized saliency maps (PSMs) reflect the gaze patterns of different subjects. Current works of saliency prediction explore the common trend in fixation distribution across all observers as a task. Personalized saliency prediction takes the personal preference of each individual into account. In other words, it considers each subject as a task. Due to the difficulty of obtaining individual labeled data, limited works are focusing on personalized saliency prediction. Our goal is to train a model that can quickly adapt to a new subject by using a few labeled data from the new subject. This paper proposes a meta-learning-based method to solve the few-shot personalized saliency prediction problem and predict better saliency maps of different subjects. Our method learns model parameters to fast adapt to new subjects. In addition, we design a Hard Samples Selection (HSS) strategy to make the training process more effective. Specifically, due to the adaptability of the model to different subjects is reflected in the value of the loss function during training, we regard subjects with high loss function values as hard samples. Then we can select hard samples online and retrain the model based on them to improve the saliency prediction performance. Experimental results show that our proposed method is better than existing methods on the PSM dataset for personalized saliency prediction.