visual label
Reviews: Partially-Supervised Image Captioning
The authors tackle the problem of image captioning in the setting where captions may not be available for novel' categories. This setting is well motivated by the authors and is a practical one - it is easier to scale image-level annotations than it is to get image captions. They present an approach that uses some paired image captions and visual labels of novel categories to caption novel categories. The visual labels of novel categories are combined using an automata to only accept captions that contain some mention of the new classes. The method is evaluated on the novel class split proposed by Hendricks et al.