Abstract
Recently, there has been much interest in developing image captioning models. State-of-the-art models reached a good performance in producing human-like descriptions from image features that are extracted from neural network models such as CNN and R-CNN. However, none of the previous methods have encapsulated explicit features that reflect a human perception of the images such as gazing patterns without the use of the eye-tracking systems. In this paper, we hypothesize that the nouns (i.e. entities) and their orders in the image description reflect human gazing patterns and perception. To this end, we estimate the sequence of the gazed objects from the words in the captions and then train a pointer network to learn to produce such sequence automatically given a set of objects in new images. We incorporate the suggested sequence by pointer network in existing image caption models and investigate its performance. Our experiments show a significant increase in the performance of the image captioning models when the sequence of the gazed objects are utilized as additional features (up to 13 points improvement in CIDEr score when combined with Neural Image Caption model).