Bridging Modality Gap for Visual Grounding with Effecitve Cross-modal Distillation

Visual grounding aims to align visual information of specific regions of images with corresponding natural language expressions. Current visual grounding methods leverage pre-trained visual and language backbones separately to obtain visual features and linguistic features. Although these two types of features are then fused via delicately designed networks, the heterogeneity of the features makes them inapplicable for multi-modal reasoning. This problem arises from the domain gap between the single-modal pre-training backbone used in current visual grounding methods, which can hardly be overcome by the traditional end-to-end training method. To alleviate this, our work proposes an Empowering pre-trained model for Visual Grounding (EpmVG) framework, which distills a multimodal pre-trained model to guide the visual grounding task. EpmVG is based on a novel cross-modal distillation mechanism, which can effectively introduce the consistency information of images and texts in the pre-trained model, to reduce the domain gap existing in the backbone networks, thereby improving the performance of the model in the visual grounding task. Extensive experiments are carried out on five conventionally used datasets, and results demonstrate that our method achieves better performance than state-of-the-art methods.