MMT: A Multimodal Translator for Image Captioning

Image captioning is a challenging problem. Different from other computer vision tasks such as image classification and object detection, image captioning requires not only understanding the image, but also the knowledge of natural language. In this work, we formulate the problem of image captioning as a multimodal translation task. Analogous to machine translation, we present a sequence-to-sequence recurrent neural network (RNN) model for image caption generation. Different from most existing work where the whole image is represented by a convolutional neural network (CNN) feature, we propose to represent the input image as a sequence of detected objects to serve as the source sequence of the RNN model. In this way, the sequential representation of an image can be naturally translated into a sequence of words, as the target sequence of the RNN model. To obtain the source sequence from the image, objects are first detected by pre-trained detectors and then converted to a sequential representation using heuristic ordering strategies, that is, by the saliency scores of the detected objects. We propose three ordering methods, descending, ascending and random, according to the saliency scores, in order to study the influence of ordering over RNN cells. To obtain the target sequence, the language words are represented as one-hot feature vector. The representations of the objects and the words are then mapped into a common hidden space. The translation from the source sequence to the target sequence is done by leveraging LSTM. Extensive experiments are conducted to evaluate the proposed approach on benchmark dataset, i.e., MSCOCO, and achieve the state-of-the-art performance. The proposed approach is also evaluated by the evaluation server of MS COCO captioning challenge and achieves very competitive results. For example, we achieve CIDEr of 93.2, RougeL of 53.2 and BLEU4 of 31.1. We validate the contribution of each idea, that is, sequential representation and ordering method, by comparison studies, and show that sequential representation indeed improves the performance compared to vanilla CNN + RNN based methods, and ascending ordering outperforms the other two ordering methods.