Target-Absent Human Attention

AbstractThe prediction of human gaze behavior is important for building human-computer interaction systems that can anticipate the user’s attention. Computer vision models have been developed to predict fixations made by people as they search target objects. But what about when not in image? Equally know how cannot find a target, and would stop searching. In this paper, we propose data-driven computational model addresses search-termination problem predicts scanpath searching targets do appear images. We visual an imitation learning represent internal knowledge viewer acquires through using novel state representation call Foveated Feature Maps (FFMs). FFMs integrate simulated foveated retina into pretrained ConvNet produces in-network feature pyramid, all with minimal overhead. Our method integrates inverse reinforcement learning. Experimentally, improve art predicting target-absent on COCO-Search18 dataset. Code available at: https://github.com/cvlab-stonybrook/Target-absent-Human-Attention.KeywordsVisual searchHuman attentionInverse learningScanpath predictionTermination predictionTarget absent