Harlequin: Color-driven Generation of Synthetic Data for Referring Expression Comprehension

Referring Expression Comprehension (REC) aims to identify a particular object in a scene by a natural language expression, and is an important topic in visual language understanding. State-of-the-art methods for this task are based on deep learning, which generally requires expensive and manually labeled annotations. Some works tackle the problem with limited-supervision learning or relying on Large Vision and Language Models. However, the development of techniques to synthesize labeled data is overlooked. In this paper, we propose a novel framework that generates artificial data for the REC task, taking into account both textual and visual modalities. At first, our pipeline processes existing data to create variations in the annotations. Then, it generates an image using altered annotations as guidance. The result of this pipeline is a new dataset, called Harlequin, made by more than 1M queries. This approach eliminates manual data collection and annotation, enabling scalability and facilitating arbitrary complexity. We pre-train three REC models on Harlequin, then fine-tuned and evaluated on human-annotated datasets. Our experiments show that the pre-training on artificial data is beneficial for performance.

Weakly-Supervised Visual-Textual Grounding with Semantic Prior Refinement

Using only image-sentence pairs, weakly-supervised visual-textual grounding aims to learn region-phrase correspondences of the respective entity mentions. Compared to the supervised approach, learning is more difficult since bounding boxes and textual phrases correspondences are unavailable. In light of this, we propose the Semantic Prior Refinement Model (SPRM), whose predictions are obtained by combining the output of two main modules. The first untrained module aims to return a rough alignment between textual phrases and bounding boxes. The second trained module is composed of two sub-components that refine the rough alignment to improve the accuracy of the final phrase-bounding box alignments. The model is trained to maximize the multimodal similarity between an image and a sentence, while minimizing the multimodal similarity of the same sentence and a new unrelated image, carefully selected to help the most during training. Our approach shows state-of-the-art results on two popular datasets, Flickr30k Entities and ReferIt, shining especially on ReferIt with a 9.6% absolute improvement. Moreover, thanks to the untrained component, it reaches competitive performances just using a small fraction of training examples.