Saliency Map Extraction in Human Crowd RGB Data

Abstract

Saliency map in human crowded scene is a prediction of regions which attracts human visual attention. Humans have an ability to analyze the context of visual scene and focus their attention to salient regions in the crowd scene. In this work, we propose a novel convolutional neural network based method for saliency prediction. Unlike classical works on crowd scene using hand-crafted face features, our model extracts deep features using convolutional layers from image classification model and learns the global context using large receptive convolutional layers. Self-attention mechanism is applied to detect the dependency between elements of feature maps. This model overperformed state-of-the-art methods on the saliency in human crowd Eyecrowd dataset.