Grounded Image Captioning in Top-down View

Weakly supervised grounded image captioning (WSGIC) aims to generate the caption and ground (localize) predicted object words in the input image without using bounding box supervision. Recent two-stage solutions mostly apply a bottom-up pipeline: (1) first apply an off-the-shelf object detector to encode the input image into multiple region features; (2) and then leverage a soft-attention mechanism for captioning and grounding. However, object detectors are mainly designed to extract object semantics (i.e., the object category). Besides, they break down the structural images into pieces of individual proposals. As a result, the subsequent grounded captioner is often overfitted to find the correct object words, while overlooking the relation between objects (e.g., what is the person doing?), and selecting incompatible proposal regions for grounding. To address these difficulties, we propose a one-stage weakly supervised grounded captioner that directly takes the RGB image as input to perform captioning and grounding at the top-down image level. In addition, we explicitly inject a relation module into our one-stage framework to encourage the relation understanding through multi-label classification. The relation semantics aid the prediction of relation words in the caption. We observe that the relation words not only assist the grounded captioner in generating a more accurate caption but also improve the grounding performance. We validate the effectiveness of our proposed method on two challenging datasets (Flick30k Entities captioning and MSCOCO captioning). The experimental results demonstrate that our method achieves state-of-the-art grounding performance.