Learning to focus on target for weakly supervised visual grounding

Zhi Xu, Yun Fu

Keywords: Visual grounding, Vision-language model

Abstract: Visual grounding is a crucial task for connecting visual and language descriptions by identifying target objects based on language entities. However, fully supervised methods require extensive annotations, which can be challenging and time-consuming. Weakly supervised visual grounding, which only relies on image-sentence association without object-level annotations, offers a promising solution. Previous approaches have mainly focused on finding the relationship between detected candidates, without considering improving object localization. In this work, we propose a novel method that leverages Grad-CAM to help the model identify precise objects. Specifically, we introduce a CAM encoder that exploits Grad-CAM information and a new loss function, attention mining loss, to guide the Grad-CAM feature to focus on the entire object. We also use an architecture that combines CNN and transformer, and a multi-modality fusion module to aggregate visual features, language features, and CAM features. Our proposed approach achieves state-of-the-art results on several datasets, demonstrating its effectiveness in different scenes. Ablation studies further confirm the benefits of our architecture.

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3706